Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * mm/percpu.c - percpu memory allocator
4 *
5 * Copyright (C) 2009 SUSE Linux Products GmbH
6 * Copyright (C) 2009 Tejun Heo <tj@kernel.org>
7 *
8 * Copyright (C) 2017 Facebook Inc.
9 * Copyright (C) 2017 Dennis Zhou <dennis@kernel.org>
10 *
11 * The percpu allocator handles both static and dynamic areas. Percpu
12 * areas are allocated in chunks which are divided into units. There is
13 * a 1-to-1 mapping for units to possible cpus. These units are grouped
14 * based on NUMA properties of the machine.
15 *
16 * c0 c1 c2
17 * ------------------- ------------------- ------------
18 * | u0 | u1 | u2 | u3 | | u0 | u1 | u2 | u3 | | u0 | u1 | u
19 * ------------------- ...... ------------------- .... ------------
20 *
21 * Allocation is done by offsets into a unit's address space. Ie., an
22 * area of 512 bytes at 6k in c1 occupies 512 bytes at 6k in c1:u0,
23 * c1:u1, c1:u2, etc. On NUMA machines, the mapping may be non-linear
24 * and even sparse. Access is handled by configuring percpu base
25 * registers according to the cpu to unit mappings and offsetting the
26 * base address using pcpu_unit_size.
27 *
28 * There is special consideration for the first chunk which must handle
29 * the static percpu variables in the kernel image as allocation services
30 * are not online yet. In short, the first chunk is structured like so:
31 *
32 * <Static | [Reserved] | Dynamic>
33 *
34 * The static data is copied from the original section managed by the
35 * linker. The reserved section, if non-zero, primarily manages static
36 * percpu variables from kernel modules. Finally, the dynamic section
37 * takes care of normal allocations.
38 *
39 * The allocator organizes chunks into lists according to free size and
40 * memcg-awareness. To make a percpu allocation memcg-aware the __GFP_ACCOUNT
41 * flag should be passed. All memcg-aware allocations are sharing one set
42 * of chunks and all unaccounted allocations and allocations performed
43 * by processes belonging to the root memory cgroup are using the second set.
44 *
45 * The allocator tries to allocate from the fullest chunk first. Each chunk
46 * is managed by a bitmap with metadata blocks. The allocation map is updated
47 * on every allocation and free to reflect the current state while the boundary
48 * map is only updated on allocation. Each metadata block contains
49 * information to help mitigate the need to iterate over large portions
50 * of the bitmap. The reverse mapping from page to chunk is stored in
51 * the page's index. Lastly, units are lazily backed and grow in unison.
52 *
53 * There is a unique conversion that goes on here between bytes and bits.
54 * Each bit represents a fragment of size PCPU_MIN_ALLOC_SIZE. The chunk
55 * tracks the number of pages it is responsible for in nr_pages. Helper
56 * functions are used to convert from between the bytes, bits, and blocks.
57 * All hints are managed in bits unless explicitly stated.
58 *
59 * To use this allocator, arch code should do the following:
60 *
61 * - define __addr_to_pcpu_ptr() and __pcpu_ptr_to_addr() to translate
62 * regular address to percpu pointer and back if they need to be
63 * different from the default
64 *
65 * - use pcpu_setup_first_chunk() during percpu area initialization to
66 * setup the first chunk containing the kernel static percpu area
67 */
68
69#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
70
71#include <linux/bitmap.h>
72#include <linux/cpumask.h>
73#include <linux/memblock.h>
74#include <linux/err.h>
75#include <linux/list.h>
76#include <linux/log2.h>
77#include <linux/mm.h>
78#include <linux/module.h>
79#include <linux/mutex.h>
80#include <linux/percpu.h>
81#include <linux/pfn.h>
82#include <linux/slab.h>
83#include <linux/spinlock.h>
84#include <linux/vmalloc.h>
85#include <linux/workqueue.h>
86#include <linux/kmemleak.h>
87#include <linux/sched.h>
88#include <linux/sched/mm.h>
89#include <linux/memcontrol.h>
90
91#include <asm/cacheflush.h>
92#include <asm/sections.h>
93#include <asm/tlbflush.h>
94#include <asm/io.h>
95
96#define CREATE_TRACE_POINTS
97#include <trace/events/percpu.h>
98
99#include "percpu-internal.h"
100
101/*
102 * The slots are sorted by the size of the biggest continuous free area.
103 * 1-31 bytes share the same slot.
104 */
105#define PCPU_SLOT_BASE_SHIFT 5
106/* chunks in slots below this are subject to being sidelined on failed alloc */
107#define PCPU_SLOT_FAIL_THRESHOLD 3
108
109#define PCPU_EMPTY_POP_PAGES_LOW 2
110#define PCPU_EMPTY_POP_PAGES_HIGH 4
111
112#ifdef CONFIG_SMP
113/* default addr <-> pcpu_ptr mapping, override in asm/percpu.h if necessary */
114#ifndef __addr_to_pcpu_ptr
115#define __addr_to_pcpu_ptr(addr) \
116 (void __percpu *)((unsigned long)(addr) - \
117 (unsigned long)pcpu_base_addr + \
118 (unsigned long)__per_cpu_start)
119#endif
120#ifndef __pcpu_ptr_to_addr
121#define __pcpu_ptr_to_addr(ptr) \
122 (void __force *)((unsigned long)(ptr) + \
123 (unsigned long)pcpu_base_addr - \
124 (unsigned long)__per_cpu_start)
125#endif
126#else /* CONFIG_SMP */
127/* on UP, it's always identity mapped */
128#define __addr_to_pcpu_ptr(addr) (void __percpu *)(addr)
129#define __pcpu_ptr_to_addr(ptr) (void __force *)(ptr)
130#endif /* CONFIG_SMP */
131
132static int pcpu_unit_pages __ro_after_init;
133static int pcpu_unit_size __ro_after_init;
134static int pcpu_nr_units __ro_after_init;
135static int pcpu_atom_size __ro_after_init;
136int pcpu_nr_slots __ro_after_init;
137static int pcpu_free_slot __ro_after_init;
138int pcpu_sidelined_slot __ro_after_init;
139int pcpu_to_depopulate_slot __ro_after_init;
140static size_t pcpu_chunk_struct_size __ro_after_init;
141
142/* cpus with the lowest and highest unit addresses */
143static unsigned int pcpu_low_unit_cpu __ro_after_init;
144static unsigned int pcpu_high_unit_cpu __ro_after_init;
145
146/* the address of the first chunk which starts with the kernel static area */
147void *pcpu_base_addr __ro_after_init;
148
149static const int *pcpu_unit_map __ro_after_init; /* cpu -> unit */
150const unsigned long *pcpu_unit_offsets __ro_after_init; /* cpu -> unit offset */
151
152/* group information, used for vm allocation */
153static int pcpu_nr_groups __ro_after_init;
154static const unsigned long *pcpu_group_offsets __ro_after_init;
155static const size_t *pcpu_group_sizes __ro_after_init;
156
157/*
158 * The first chunk which always exists. Note that unlike other
159 * chunks, this one can be allocated and mapped in several different
160 * ways and thus often doesn't live in the vmalloc area.
161 */
162struct pcpu_chunk *pcpu_first_chunk __ro_after_init;
163
164/*
165 * Optional reserved chunk. This chunk reserves part of the first
166 * chunk and serves it for reserved allocations. When the reserved
167 * region doesn't exist, the following variable is NULL.
168 */
169struct pcpu_chunk *pcpu_reserved_chunk __ro_after_init;
170
171DEFINE_SPINLOCK(pcpu_lock); /* all internal data structures */
172static DEFINE_MUTEX(pcpu_alloc_mutex); /* chunk create/destroy, [de]pop, map ext */
173
174struct list_head *pcpu_chunk_lists __ro_after_init; /* chunk list slots */
175
176/*
177 * The number of empty populated pages, protected by pcpu_lock.
178 * The reserved chunk doesn't contribute to the count.
179 */
180int pcpu_nr_empty_pop_pages;
181
182/*
183 * The number of populated pages in use by the allocator, protected by
184 * pcpu_lock. This number is kept per a unit per chunk (i.e. when a page gets
185 * allocated/deallocated, it is allocated/deallocated in all units of a chunk
186 * and increments/decrements this count by 1).
187 */
188static unsigned long pcpu_nr_populated;
189
190/*
191 * Balance work is used to populate or destroy chunks asynchronously. We
192 * try to keep the number of populated free pages between
193 * PCPU_EMPTY_POP_PAGES_LOW and HIGH for atomic allocations and at most one
194 * empty chunk.
195 */
196static void pcpu_balance_workfn(struct work_struct *work);
197static DECLARE_WORK(pcpu_balance_work, pcpu_balance_workfn);
198static bool pcpu_async_enabled __read_mostly;
199static bool pcpu_atomic_alloc_failed;
200
201static void pcpu_schedule_balance_work(void)
202{
203 if (pcpu_async_enabled)
204 schedule_work(&pcpu_balance_work);
205}
206
207/**
208 * pcpu_addr_in_chunk - check if the address is served from this chunk
209 * @chunk: chunk of interest
210 * @addr: percpu address
211 *
212 * RETURNS:
213 * True if the address is served from this chunk.
214 */
215static bool pcpu_addr_in_chunk(struct pcpu_chunk *chunk, void *addr)
216{
217 void *start_addr, *end_addr;
218
219 if (!chunk)
220 return false;
221
222 start_addr = chunk->base_addr + chunk->start_offset;
223 end_addr = chunk->base_addr + chunk->nr_pages * PAGE_SIZE -
224 chunk->end_offset;
225
226 return addr >= start_addr && addr < end_addr;
227}
228
229static int __pcpu_size_to_slot(int size)
230{
231 int highbit = fls(size); /* size is in bytes */
232 return max(highbit - PCPU_SLOT_BASE_SHIFT + 2, 1);
233}
234
235static int pcpu_size_to_slot(int size)
236{
237 if (size == pcpu_unit_size)
238 return pcpu_free_slot;
239 return __pcpu_size_to_slot(size);
240}
241
242static int pcpu_chunk_slot(const struct pcpu_chunk *chunk)
243{
244 const struct pcpu_block_md *chunk_md = &chunk->chunk_md;
245
246 if (chunk->free_bytes < PCPU_MIN_ALLOC_SIZE ||
247 chunk_md->contig_hint == 0)
248 return 0;
249
250 return pcpu_size_to_slot(chunk_md->contig_hint * PCPU_MIN_ALLOC_SIZE);
251}
252
253/* set the pointer to a chunk in a page struct */
254static void pcpu_set_page_chunk(struct page *page, struct pcpu_chunk *pcpu)
255{
256 page->private = (unsigned long)pcpu;
257}
258
259/* obtain pointer to a chunk from a page struct */
260static struct pcpu_chunk *pcpu_get_page_chunk(struct page *page)
261{
262 return (struct pcpu_chunk *)page->private;
263}
264
265static int __maybe_unused pcpu_page_idx(unsigned int cpu, int page_idx)
266{
267 return pcpu_unit_map[cpu] * pcpu_unit_pages + page_idx;
268}
269
270static unsigned long pcpu_unit_page_offset(unsigned int cpu, int page_idx)
271{
272 return pcpu_unit_offsets[cpu] + (page_idx << PAGE_SHIFT);
273}
274
275static unsigned long pcpu_chunk_addr(struct pcpu_chunk *chunk,
276 unsigned int cpu, int page_idx)
277{
278 return (unsigned long)chunk->base_addr +
279 pcpu_unit_page_offset(cpu, page_idx);
280}
281
282/*
283 * The following are helper functions to help access bitmaps and convert
284 * between bitmap offsets to address offsets.
285 */
286static unsigned long *pcpu_index_alloc_map(struct pcpu_chunk *chunk, int index)
287{
288 return chunk->alloc_map +
289 (index * PCPU_BITMAP_BLOCK_BITS / BITS_PER_LONG);
290}
291
292static unsigned long pcpu_off_to_block_index(int off)
293{
294 return off / PCPU_BITMAP_BLOCK_BITS;
295}
296
297static unsigned long pcpu_off_to_block_off(int off)
298{
299 return off & (PCPU_BITMAP_BLOCK_BITS - 1);
300}
301
302static unsigned long pcpu_block_off_to_off(int index, int off)
303{
304 return index * PCPU_BITMAP_BLOCK_BITS + off;
305}
306
307/**
308 * pcpu_check_block_hint - check against the contig hint
309 * @block: block of interest
310 * @bits: size of allocation
311 * @align: alignment of area (max PAGE_SIZE)
312 *
313 * Check to see if the allocation can fit in the block's contig hint.
314 * Note, a chunk uses the same hints as a block so this can also check against
315 * the chunk's contig hint.
316 */
317static bool pcpu_check_block_hint(struct pcpu_block_md *block, int bits,
318 size_t align)
319{
320 int bit_off = ALIGN(block->contig_hint_start, align) -
321 block->contig_hint_start;
322
323 return bit_off + bits <= block->contig_hint;
324}
325
326/*
327 * pcpu_next_hint - determine which hint to use
328 * @block: block of interest
329 * @alloc_bits: size of allocation
330 *
331 * This determines if we should scan based on the scan_hint or first_free.
332 * In general, we want to scan from first_free to fulfill allocations by
333 * first fit. However, if we know a scan_hint at position scan_hint_start
334 * cannot fulfill an allocation, we can begin scanning from there knowing
335 * the contig_hint will be our fallback.
336 */
337static int pcpu_next_hint(struct pcpu_block_md *block, int alloc_bits)
338{
339 /*
340 * The three conditions below determine if we can skip past the
341 * scan_hint. First, does the scan hint exist. Second, is the
342 * contig_hint after the scan_hint (possibly not true iff
343 * contig_hint == scan_hint). Third, is the allocation request
344 * larger than the scan_hint.
345 */
346 if (block->scan_hint &&
347 block->contig_hint_start > block->scan_hint_start &&
348 alloc_bits > block->scan_hint)
349 return block->scan_hint_start + block->scan_hint;
350
351 return block->first_free;
352}
353
354/**
355 * pcpu_next_md_free_region - finds the next hint free area
356 * @chunk: chunk of interest
357 * @bit_off: chunk offset
358 * @bits: size of free area
359 *
360 * Helper function for pcpu_for_each_md_free_region. It checks
361 * block->contig_hint and performs aggregation across blocks to find the
362 * next hint. It modifies bit_off and bits in-place to be consumed in the
363 * loop.
364 */
365static void pcpu_next_md_free_region(struct pcpu_chunk *chunk, int *bit_off,
366 int *bits)
367{
368 int i = pcpu_off_to_block_index(*bit_off);
369 int block_off = pcpu_off_to_block_off(*bit_off);
370 struct pcpu_block_md *block;
371
372 *bits = 0;
373 for (block = chunk->md_blocks + i; i < pcpu_chunk_nr_blocks(chunk);
374 block++, i++) {
375 /* handles contig area across blocks */
376 if (*bits) {
377 *bits += block->left_free;
378 if (block->left_free == PCPU_BITMAP_BLOCK_BITS)
379 continue;
380 return;
381 }
382
383 /*
384 * This checks three things. First is there a contig_hint to
385 * check. Second, have we checked this hint before by
386 * comparing the block_off. Third, is this the same as the
387 * right contig hint. In the last case, it spills over into
388 * the next block and should be handled by the contig area
389 * across blocks code.
390 */
391 *bits = block->contig_hint;
392 if (*bits && block->contig_hint_start >= block_off &&
393 *bits + block->contig_hint_start < PCPU_BITMAP_BLOCK_BITS) {
394 *bit_off = pcpu_block_off_to_off(i,
395 block->contig_hint_start);
396 return;
397 }
398 /* reset to satisfy the second predicate above */
399 block_off = 0;
400
401 *bits = block->right_free;
402 *bit_off = (i + 1) * PCPU_BITMAP_BLOCK_BITS - block->right_free;
403 }
404}
405
406/**
407 * pcpu_next_fit_region - finds fit areas for a given allocation request
408 * @chunk: chunk of interest
409 * @alloc_bits: size of allocation
410 * @align: alignment of area (max PAGE_SIZE)
411 * @bit_off: chunk offset
412 * @bits: size of free area
413 *
414 * Finds the next free region that is viable for use with a given size and
415 * alignment. This only returns if there is a valid area to be used for this
416 * allocation. block->first_free is returned if the allocation request fits
417 * within the block to see if the request can be fulfilled prior to the contig
418 * hint.
419 */
420static void pcpu_next_fit_region(struct pcpu_chunk *chunk, int alloc_bits,
421 int align, int *bit_off, int *bits)
422{
423 int i = pcpu_off_to_block_index(*bit_off);
424 int block_off = pcpu_off_to_block_off(*bit_off);
425 struct pcpu_block_md *block;
426
427 *bits = 0;
428 for (block = chunk->md_blocks + i; i < pcpu_chunk_nr_blocks(chunk);
429 block++, i++) {
430 /* handles contig area across blocks */
431 if (*bits) {
432 *bits += block->left_free;
433 if (*bits >= alloc_bits)
434 return;
435 if (block->left_free == PCPU_BITMAP_BLOCK_BITS)
436 continue;
437 }
438
439 /* check block->contig_hint */
440 *bits = ALIGN(block->contig_hint_start, align) -
441 block->contig_hint_start;
442 /*
443 * This uses the block offset to determine if this has been
444 * checked in the prior iteration.
445 */
446 if (block->contig_hint &&
447 block->contig_hint_start >= block_off &&
448 block->contig_hint >= *bits + alloc_bits) {
449 int start = pcpu_next_hint(block, alloc_bits);
450
451 *bits += alloc_bits + block->contig_hint_start -
452 start;
453 *bit_off = pcpu_block_off_to_off(i, start);
454 return;
455 }
456 /* reset to satisfy the second predicate above */
457 block_off = 0;
458
459 *bit_off = ALIGN(PCPU_BITMAP_BLOCK_BITS - block->right_free,
460 align);
461 *bits = PCPU_BITMAP_BLOCK_BITS - *bit_off;
462 *bit_off = pcpu_block_off_to_off(i, *bit_off);
463 if (*bits >= alloc_bits)
464 return;
465 }
466
467 /* no valid offsets were found - fail condition */
468 *bit_off = pcpu_chunk_map_bits(chunk);
469}
470
471/*
472 * Metadata free area iterators. These perform aggregation of free areas
473 * based on the metadata blocks and return the offset @bit_off and size in
474 * bits of the free area @bits. pcpu_for_each_fit_region only returns when
475 * a fit is found for the allocation request.
476 */
477#define pcpu_for_each_md_free_region(chunk, bit_off, bits) \
478 for (pcpu_next_md_free_region((chunk), &(bit_off), &(bits)); \
479 (bit_off) < pcpu_chunk_map_bits((chunk)); \
480 (bit_off) += (bits) + 1, \
481 pcpu_next_md_free_region((chunk), &(bit_off), &(bits)))
482
483#define pcpu_for_each_fit_region(chunk, alloc_bits, align, bit_off, bits) \
484 for (pcpu_next_fit_region((chunk), (alloc_bits), (align), &(bit_off), \
485 &(bits)); \
486 (bit_off) < pcpu_chunk_map_bits((chunk)); \
487 (bit_off) += (bits), \
488 pcpu_next_fit_region((chunk), (alloc_bits), (align), &(bit_off), \
489 &(bits)))
490
491/**
492 * pcpu_mem_zalloc - allocate memory
493 * @size: bytes to allocate
494 * @gfp: allocation flags
495 *
496 * Allocate @size bytes. If @size is smaller than PAGE_SIZE,
497 * kzalloc() is used; otherwise, the equivalent of vzalloc() is used.
498 * This is to facilitate passing through whitelisted flags. The
499 * returned memory is always zeroed.
500 *
501 * RETURNS:
502 * Pointer to the allocated area on success, NULL on failure.
503 */
504static void *pcpu_mem_zalloc(size_t size, gfp_t gfp)
505{
506 if (WARN_ON_ONCE(!slab_is_available()))
507 return NULL;
508
509 if (size <= PAGE_SIZE)
510 return kzalloc(size, gfp);
511 else
512 return __vmalloc(size, gfp | __GFP_ZERO);
513}
514
515/**
516 * pcpu_mem_free - free memory
517 * @ptr: memory to free
518 *
519 * Free @ptr. @ptr should have been allocated using pcpu_mem_zalloc().
520 */
521static void pcpu_mem_free(void *ptr)
522{
523 kvfree(ptr);
524}
525
526static void __pcpu_chunk_move(struct pcpu_chunk *chunk, int slot,
527 bool move_front)
528{
529 if (chunk != pcpu_reserved_chunk) {
530 if (move_front)
531 list_move(&chunk->list, &pcpu_chunk_lists[slot]);
532 else
533 list_move_tail(&chunk->list, &pcpu_chunk_lists[slot]);
534 }
535}
536
537static void pcpu_chunk_move(struct pcpu_chunk *chunk, int slot)
538{
539 __pcpu_chunk_move(chunk, slot, true);
540}
541
542/**
543 * pcpu_chunk_relocate - put chunk in the appropriate chunk slot
544 * @chunk: chunk of interest
545 * @oslot: the previous slot it was on
546 *
547 * This function is called after an allocation or free changed @chunk.
548 * New slot according to the changed state is determined and @chunk is
549 * moved to the slot. Note that the reserved chunk is never put on
550 * chunk slots.
551 *
552 * CONTEXT:
553 * pcpu_lock.
554 */
555static void pcpu_chunk_relocate(struct pcpu_chunk *chunk, int oslot)
556{
557 int nslot = pcpu_chunk_slot(chunk);
558
559 /* leave isolated chunks in-place */
560 if (chunk->isolated)
561 return;
562
563 if (oslot != nslot)
564 __pcpu_chunk_move(chunk, nslot, oslot < nslot);
565}
566
567static void pcpu_isolate_chunk(struct pcpu_chunk *chunk)
568{
569 lockdep_assert_held(&pcpu_lock);
570
571 if (!chunk->isolated) {
572 chunk->isolated = true;
573 pcpu_nr_empty_pop_pages -= chunk->nr_empty_pop_pages;
574 }
575 list_move(&chunk->list, &pcpu_chunk_lists[pcpu_to_depopulate_slot]);
576}
577
578static void pcpu_reintegrate_chunk(struct pcpu_chunk *chunk)
579{
580 lockdep_assert_held(&pcpu_lock);
581
582 if (chunk->isolated) {
583 chunk->isolated = false;
584 pcpu_nr_empty_pop_pages += chunk->nr_empty_pop_pages;
585 pcpu_chunk_relocate(chunk, -1);
586 }
587}
588
589/*
590 * pcpu_update_empty_pages - update empty page counters
591 * @chunk: chunk of interest
592 * @nr: nr of empty pages
593 *
594 * This is used to keep track of the empty pages now based on the premise
595 * a md_block covers a page. The hint update functions recognize if a block
596 * is made full or broken to calculate deltas for keeping track of free pages.
597 */
598static inline void pcpu_update_empty_pages(struct pcpu_chunk *chunk, int nr)
599{
600 chunk->nr_empty_pop_pages += nr;
601 if (chunk != pcpu_reserved_chunk && !chunk->isolated)
602 pcpu_nr_empty_pop_pages += nr;
603}
604
605/*
606 * pcpu_region_overlap - determines if two regions overlap
607 * @a: start of first region, inclusive
608 * @b: end of first region, exclusive
609 * @x: start of second region, inclusive
610 * @y: end of second region, exclusive
611 *
612 * This is used to determine if the hint region [a, b) overlaps with the
613 * allocated region [x, y).
614 */
615static inline bool pcpu_region_overlap(int a, int b, int x, int y)
616{
617 return (a < y) && (x < b);
618}
619
620/**
621 * pcpu_block_update - updates a block given a free area
622 * @block: block of interest
623 * @start: start offset in block
624 * @end: end offset in block
625 *
626 * Updates a block given a known free area. The region [start, end) is
627 * expected to be the entirety of the free area within a block. Chooses
628 * the best starting offset if the contig hints are equal.
629 */
630static void pcpu_block_update(struct pcpu_block_md *block, int start, int end)
631{
632 int contig = end - start;
633
634 block->first_free = min(block->first_free, start);
635 if (start == 0)
636 block->left_free = contig;
637
638 if (end == block->nr_bits)
639 block->right_free = contig;
640
641 if (contig > block->contig_hint) {
642 /* promote the old contig_hint to be the new scan_hint */
643 if (start > block->contig_hint_start) {
644 if (block->contig_hint > block->scan_hint) {
645 block->scan_hint_start =
646 block->contig_hint_start;
647 block->scan_hint = block->contig_hint;
648 } else if (start < block->scan_hint_start) {
649 /*
650 * The old contig_hint == scan_hint. But, the
651 * new contig is larger so hold the invariant
652 * scan_hint_start < contig_hint_start.
653 */
654 block->scan_hint = 0;
655 }
656 } else {
657 block->scan_hint = 0;
658 }
659 block->contig_hint_start = start;
660 block->contig_hint = contig;
661 } else if (contig == block->contig_hint) {
662 if (block->contig_hint_start &&
663 (!start ||
664 __ffs(start) > __ffs(block->contig_hint_start))) {
665 /* start has a better alignment so use it */
666 block->contig_hint_start = start;
667 if (start < block->scan_hint_start &&
668 block->contig_hint > block->scan_hint)
669 block->scan_hint = 0;
670 } else if (start > block->scan_hint_start ||
671 block->contig_hint > block->scan_hint) {
672 /*
673 * Knowing contig == contig_hint, update the scan_hint
674 * if it is farther than or larger than the current
675 * scan_hint.
676 */
677 block->scan_hint_start = start;
678 block->scan_hint = contig;
679 }
680 } else {
681 /*
682 * The region is smaller than the contig_hint. So only update
683 * the scan_hint if it is larger than or equal and farther than
684 * the current scan_hint.
685 */
686 if ((start < block->contig_hint_start &&
687 (contig > block->scan_hint ||
688 (contig == block->scan_hint &&
689 start > block->scan_hint_start)))) {
690 block->scan_hint_start = start;
691 block->scan_hint = contig;
692 }
693 }
694}
695
696/*
697 * pcpu_block_update_scan - update a block given a free area from a scan
698 * @chunk: chunk of interest
699 * @bit_off: chunk offset
700 * @bits: size of free area
701 *
702 * Finding the final allocation spot first goes through pcpu_find_block_fit()
703 * to find a block that can hold the allocation and then pcpu_alloc_area()
704 * where a scan is used. When allocations require specific alignments,
705 * we can inadvertently create holes which will not be seen in the alloc
706 * or free paths.
707 *
708 * This takes a given free area hole and updates a block as it may change the
709 * scan_hint. We need to scan backwards to ensure we don't miss free bits
710 * from alignment.
711 */
712static void pcpu_block_update_scan(struct pcpu_chunk *chunk, int bit_off,
713 int bits)
714{
715 int s_off = pcpu_off_to_block_off(bit_off);
716 int e_off = s_off + bits;
717 int s_index, l_bit;
718 struct pcpu_block_md *block;
719
720 if (e_off > PCPU_BITMAP_BLOCK_BITS)
721 return;
722
723 s_index = pcpu_off_to_block_index(bit_off);
724 block = chunk->md_blocks + s_index;
725
726 /* scan backwards in case of alignment skipping free bits */
727 l_bit = find_last_bit(pcpu_index_alloc_map(chunk, s_index), s_off);
728 s_off = (s_off == l_bit) ? 0 : l_bit + 1;
729
730 pcpu_block_update(block, s_off, e_off);
731}
732
733/**
734 * pcpu_chunk_refresh_hint - updates metadata about a chunk
735 * @chunk: chunk of interest
736 * @full_scan: if we should scan from the beginning
737 *
738 * Iterates over the metadata blocks to find the largest contig area.
739 * A full scan can be avoided on the allocation path as this is triggered
740 * if we broke the contig_hint. In doing so, the scan_hint will be before
741 * the contig_hint or after if the scan_hint == contig_hint. This cannot
742 * be prevented on freeing as we want to find the largest area possibly
743 * spanning blocks.
744 */
745static void pcpu_chunk_refresh_hint(struct pcpu_chunk *chunk, bool full_scan)
746{
747 struct pcpu_block_md *chunk_md = &chunk->chunk_md;
748 int bit_off, bits;
749
750 /* promote scan_hint to contig_hint */
751 if (!full_scan && chunk_md->scan_hint) {
752 bit_off = chunk_md->scan_hint_start + chunk_md->scan_hint;
753 chunk_md->contig_hint_start = chunk_md->scan_hint_start;
754 chunk_md->contig_hint = chunk_md->scan_hint;
755 chunk_md->scan_hint = 0;
756 } else {
757 bit_off = chunk_md->first_free;
758 chunk_md->contig_hint = 0;
759 }
760
761 bits = 0;
762 pcpu_for_each_md_free_region(chunk, bit_off, bits)
763 pcpu_block_update(chunk_md, bit_off, bit_off + bits);
764}
765
766/**
767 * pcpu_block_refresh_hint
768 * @chunk: chunk of interest
769 * @index: index of the metadata block
770 *
771 * Scans over the block beginning at first_free and updates the block
772 * metadata accordingly.
773 */
774static void pcpu_block_refresh_hint(struct pcpu_chunk *chunk, int index)
775{
776 struct pcpu_block_md *block = chunk->md_blocks + index;
777 unsigned long *alloc_map = pcpu_index_alloc_map(chunk, index);
778 unsigned int start, end; /* region start, region end */
779
780 /* promote scan_hint to contig_hint */
781 if (block->scan_hint) {
782 start = block->scan_hint_start + block->scan_hint;
783 block->contig_hint_start = block->scan_hint_start;
784 block->contig_hint = block->scan_hint;
785 block->scan_hint = 0;
786 } else {
787 start = block->first_free;
788 block->contig_hint = 0;
789 }
790
791 block->right_free = 0;
792
793 /* iterate over free areas and update the contig hints */
794 for_each_clear_bitrange_from(start, end, alloc_map, PCPU_BITMAP_BLOCK_BITS)
795 pcpu_block_update(block, start, end);
796}
797
798/**
799 * pcpu_block_update_hint_alloc - update hint on allocation path
800 * @chunk: chunk of interest
801 * @bit_off: chunk offset
802 * @bits: size of request
803 *
804 * Updates metadata for the allocation path. The metadata only has to be
805 * refreshed by a full scan iff the chunk's contig hint is broken. Block level
806 * scans are required if the block's contig hint is broken.
807 */
808static void pcpu_block_update_hint_alloc(struct pcpu_chunk *chunk, int bit_off,
809 int bits)
810{
811 struct pcpu_block_md *chunk_md = &chunk->chunk_md;
812 int nr_empty_pages = 0;
813 struct pcpu_block_md *s_block, *e_block, *block;
814 int s_index, e_index; /* block indexes of the freed allocation */
815 int s_off, e_off; /* block offsets of the freed allocation */
816
817 /*
818 * Calculate per block offsets.
819 * The calculation uses an inclusive range, but the resulting offsets
820 * are [start, end). e_index always points to the last block in the
821 * range.
822 */
823 s_index = pcpu_off_to_block_index(bit_off);
824 e_index = pcpu_off_to_block_index(bit_off + bits - 1);
825 s_off = pcpu_off_to_block_off(bit_off);
826 e_off = pcpu_off_to_block_off(bit_off + bits - 1) + 1;
827
828 s_block = chunk->md_blocks + s_index;
829 e_block = chunk->md_blocks + e_index;
830
831 /*
832 * Update s_block.
833 */
834 if (s_block->contig_hint == PCPU_BITMAP_BLOCK_BITS)
835 nr_empty_pages++;
836
837 /*
838 * block->first_free must be updated if the allocation takes its place.
839 * If the allocation breaks the contig_hint, a scan is required to
840 * restore this hint.
841 */
842 if (s_off == s_block->first_free)
843 s_block->first_free = find_next_zero_bit(
844 pcpu_index_alloc_map(chunk, s_index),
845 PCPU_BITMAP_BLOCK_BITS,
846 s_off + bits);
847
848 if (pcpu_region_overlap(s_block->scan_hint_start,
849 s_block->scan_hint_start + s_block->scan_hint,
850 s_off,
851 s_off + bits))
852 s_block->scan_hint = 0;
853
854 if (pcpu_region_overlap(s_block->contig_hint_start,
855 s_block->contig_hint_start +
856 s_block->contig_hint,
857 s_off,
858 s_off + bits)) {
859 /* block contig hint is broken - scan to fix it */
860 if (!s_off)
861 s_block->left_free = 0;
862 pcpu_block_refresh_hint(chunk, s_index);
863 } else {
864 /* update left and right contig manually */
865 s_block->left_free = min(s_block->left_free, s_off);
866 if (s_index == e_index)
867 s_block->right_free = min_t(int, s_block->right_free,
868 PCPU_BITMAP_BLOCK_BITS - e_off);
869 else
870 s_block->right_free = 0;
871 }
872
873 /*
874 * Update e_block.
875 */
876 if (s_index != e_index) {
877 if (e_block->contig_hint == PCPU_BITMAP_BLOCK_BITS)
878 nr_empty_pages++;
879
880 /*
881 * When the allocation is across blocks, the end is along
882 * the left part of the e_block.
883 */
884 e_block->first_free = find_next_zero_bit(
885 pcpu_index_alloc_map(chunk, e_index),
886 PCPU_BITMAP_BLOCK_BITS, e_off);
887
888 if (e_off == PCPU_BITMAP_BLOCK_BITS) {
889 /* reset the block */
890 e_block++;
891 } else {
892 if (e_off > e_block->scan_hint_start)
893 e_block->scan_hint = 0;
894
895 e_block->left_free = 0;
896 if (e_off > e_block->contig_hint_start) {
897 /* contig hint is broken - scan to fix it */
898 pcpu_block_refresh_hint(chunk, e_index);
899 } else {
900 e_block->right_free =
901 min_t(int, e_block->right_free,
902 PCPU_BITMAP_BLOCK_BITS - e_off);
903 }
904 }
905
906 /* update in-between md_blocks */
907 nr_empty_pages += (e_index - s_index - 1);
908 for (block = s_block + 1; block < e_block; block++) {
909 block->scan_hint = 0;
910 block->contig_hint = 0;
911 block->left_free = 0;
912 block->right_free = 0;
913 }
914 }
915
916 /*
917 * If the allocation is not atomic, some blocks may not be
918 * populated with pages, while we account it here. The number
919 * of pages will be added back with pcpu_chunk_populated()
920 * when populating pages.
921 */
922 if (nr_empty_pages)
923 pcpu_update_empty_pages(chunk, -nr_empty_pages);
924
925 if (pcpu_region_overlap(chunk_md->scan_hint_start,
926 chunk_md->scan_hint_start +
927 chunk_md->scan_hint,
928 bit_off,
929 bit_off + bits))
930 chunk_md->scan_hint = 0;
931
932 /*
933 * The only time a full chunk scan is required is if the chunk
934 * contig hint is broken. Otherwise, it means a smaller space
935 * was used and therefore the chunk contig hint is still correct.
936 */
937 if (pcpu_region_overlap(chunk_md->contig_hint_start,
938 chunk_md->contig_hint_start +
939 chunk_md->contig_hint,
940 bit_off,
941 bit_off + bits))
942 pcpu_chunk_refresh_hint(chunk, false);
943}
944
945/**
946 * pcpu_block_update_hint_free - updates the block hints on the free path
947 * @chunk: chunk of interest
948 * @bit_off: chunk offset
949 * @bits: size of request
950 *
951 * Updates metadata for the allocation path. This avoids a blind block
952 * refresh by making use of the block contig hints. If this fails, it scans
953 * forward and backward to determine the extent of the free area. This is
954 * capped at the boundary of blocks.
955 *
956 * A chunk update is triggered if a page becomes free, a block becomes free,
957 * or the free spans across blocks. This tradeoff is to minimize iterating
958 * over the block metadata to update chunk_md->contig_hint.
959 * chunk_md->contig_hint may be off by up to a page, but it will never be more
960 * than the available space. If the contig hint is contained in one block, it
961 * will be accurate.
962 */
963static void pcpu_block_update_hint_free(struct pcpu_chunk *chunk, int bit_off,
964 int bits)
965{
966 int nr_empty_pages = 0;
967 struct pcpu_block_md *s_block, *e_block, *block;
968 int s_index, e_index; /* block indexes of the freed allocation */
969 int s_off, e_off; /* block offsets of the freed allocation */
970 int start, end; /* start and end of the whole free area */
971
972 /*
973 * Calculate per block offsets.
974 * The calculation uses an inclusive range, but the resulting offsets
975 * are [start, end). e_index always points to the last block in the
976 * range.
977 */
978 s_index = pcpu_off_to_block_index(bit_off);
979 e_index = pcpu_off_to_block_index(bit_off + bits - 1);
980 s_off = pcpu_off_to_block_off(bit_off);
981 e_off = pcpu_off_to_block_off(bit_off + bits - 1) + 1;
982
983 s_block = chunk->md_blocks + s_index;
984 e_block = chunk->md_blocks + e_index;
985
986 /*
987 * Check if the freed area aligns with the block->contig_hint.
988 * If it does, then the scan to find the beginning/end of the
989 * larger free area can be avoided.
990 *
991 * start and end refer to beginning and end of the free area
992 * within each their respective blocks. This is not necessarily
993 * the entire free area as it may span blocks past the beginning
994 * or end of the block.
995 */
996 start = s_off;
997 if (s_off == s_block->contig_hint + s_block->contig_hint_start) {
998 start = s_block->contig_hint_start;
999 } else {
1000 /*
1001 * Scan backwards to find the extent of the free area.
1002 * find_last_bit returns the starting bit, so if the start bit
1003 * is returned, that means there was no last bit and the
1004 * remainder of the chunk is free.
1005 */
1006 int l_bit = find_last_bit(pcpu_index_alloc_map(chunk, s_index),
1007 start);
1008 start = (start == l_bit) ? 0 : l_bit + 1;
1009 }
1010
1011 end = e_off;
1012 if (e_off == e_block->contig_hint_start)
1013 end = e_block->contig_hint_start + e_block->contig_hint;
1014 else
1015 end = find_next_bit(pcpu_index_alloc_map(chunk, e_index),
1016 PCPU_BITMAP_BLOCK_BITS, end);
1017
1018 /* update s_block */
1019 e_off = (s_index == e_index) ? end : PCPU_BITMAP_BLOCK_BITS;
1020 if (!start && e_off == PCPU_BITMAP_BLOCK_BITS)
1021 nr_empty_pages++;
1022 pcpu_block_update(s_block, start, e_off);
1023
1024 /* freeing in the same block */
1025 if (s_index != e_index) {
1026 /* update e_block */
1027 if (end == PCPU_BITMAP_BLOCK_BITS)
1028 nr_empty_pages++;
1029 pcpu_block_update(e_block, 0, end);
1030
1031 /* reset md_blocks in the middle */
1032 nr_empty_pages += (e_index - s_index - 1);
1033 for (block = s_block + 1; block < e_block; block++) {
1034 block->first_free = 0;
1035 block->scan_hint = 0;
1036 block->contig_hint_start = 0;
1037 block->contig_hint = PCPU_BITMAP_BLOCK_BITS;
1038 block->left_free = PCPU_BITMAP_BLOCK_BITS;
1039 block->right_free = PCPU_BITMAP_BLOCK_BITS;
1040 }
1041 }
1042
1043 if (nr_empty_pages)
1044 pcpu_update_empty_pages(chunk, nr_empty_pages);
1045
1046 /*
1047 * Refresh chunk metadata when the free makes a block free or spans
1048 * across blocks. The contig_hint may be off by up to a page, but if
1049 * the contig_hint is contained in a block, it will be accurate with
1050 * the else condition below.
1051 */
1052 if (((end - start) >= PCPU_BITMAP_BLOCK_BITS) || s_index != e_index)
1053 pcpu_chunk_refresh_hint(chunk, true);
1054 else
1055 pcpu_block_update(&chunk->chunk_md,
1056 pcpu_block_off_to_off(s_index, start),
1057 end);
1058}
1059
1060/**
1061 * pcpu_is_populated - determines if the region is populated
1062 * @chunk: chunk of interest
1063 * @bit_off: chunk offset
1064 * @bits: size of area
1065 * @next_off: return value for the next offset to start searching
1066 *
1067 * For atomic allocations, check if the backing pages are populated.
1068 *
1069 * RETURNS:
1070 * Bool if the backing pages are populated.
1071 * next_index is to skip over unpopulated blocks in pcpu_find_block_fit.
1072 */
1073static bool pcpu_is_populated(struct pcpu_chunk *chunk, int bit_off, int bits,
1074 int *next_off)
1075{
1076 unsigned int start, end;
1077
1078 start = PFN_DOWN(bit_off * PCPU_MIN_ALLOC_SIZE);
1079 end = PFN_UP((bit_off + bits) * PCPU_MIN_ALLOC_SIZE);
1080
1081 start = find_next_zero_bit(chunk->populated, end, start);
1082 if (start >= end)
1083 return true;
1084
1085 end = find_next_bit(chunk->populated, end, start + 1);
1086
1087 *next_off = end * PAGE_SIZE / PCPU_MIN_ALLOC_SIZE;
1088 return false;
1089}
1090
1091/**
1092 * pcpu_find_block_fit - finds the block index to start searching
1093 * @chunk: chunk of interest
1094 * @alloc_bits: size of request in allocation units
1095 * @align: alignment of area (max PAGE_SIZE bytes)
1096 * @pop_only: use populated regions only
1097 *
1098 * Given a chunk and an allocation spec, find the offset to begin searching
1099 * for a free region. This iterates over the bitmap metadata blocks to
1100 * find an offset that will be guaranteed to fit the requirements. It is
1101 * not quite first fit as if the allocation does not fit in the contig hint
1102 * of a block or chunk, it is skipped. This errs on the side of caution
1103 * to prevent excess iteration. Poor alignment can cause the allocator to
1104 * skip over blocks and chunks that have valid free areas.
1105 *
1106 * RETURNS:
1107 * The offset in the bitmap to begin searching.
1108 * -1 if no offset is found.
1109 */
1110static int pcpu_find_block_fit(struct pcpu_chunk *chunk, int alloc_bits,
1111 size_t align, bool pop_only)
1112{
1113 struct pcpu_block_md *chunk_md = &chunk->chunk_md;
1114 int bit_off, bits, next_off;
1115
1116 /*
1117 * This is an optimization to prevent scanning by assuming if the
1118 * allocation cannot fit in the global hint, there is memory pressure
1119 * and creating a new chunk would happen soon.
1120 */
1121 if (!pcpu_check_block_hint(chunk_md, alloc_bits, align))
1122 return -1;
1123
1124 bit_off = pcpu_next_hint(chunk_md, alloc_bits);
1125 bits = 0;
1126 pcpu_for_each_fit_region(chunk, alloc_bits, align, bit_off, bits) {
1127 if (!pop_only || pcpu_is_populated(chunk, bit_off, bits,
1128 &next_off))
1129 break;
1130
1131 bit_off = next_off;
1132 bits = 0;
1133 }
1134
1135 if (bit_off == pcpu_chunk_map_bits(chunk))
1136 return -1;
1137
1138 return bit_off;
1139}
1140
1141/*
1142 * pcpu_find_zero_area - modified from bitmap_find_next_zero_area_off()
1143 * @map: the address to base the search on
1144 * @size: the bitmap size in bits
1145 * @start: the bitnumber to start searching at
1146 * @nr: the number of zeroed bits we're looking for
1147 * @align_mask: alignment mask for zero area
1148 * @largest_off: offset of the largest area skipped
1149 * @largest_bits: size of the largest area skipped
1150 *
1151 * The @align_mask should be one less than a power of 2.
1152 *
1153 * This is a modified version of bitmap_find_next_zero_area_off() to remember
1154 * the largest area that was skipped. This is imperfect, but in general is
1155 * good enough. The largest remembered region is the largest failed region
1156 * seen. This does not include anything we possibly skipped due to alignment.
1157 * pcpu_block_update_scan() does scan backwards to try and recover what was
1158 * lost to alignment. While this can cause scanning to miss earlier possible
1159 * free areas, smaller allocations will eventually fill those holes.
1160 */
1161static unsigned long pcpu_find_zero_area(unsigned long *map,
1162 unsigned long size,
1163 unsigned long start,
1164 unsigned long nr,
1165 unsigned long align_mask,
1166 unsigned long *largest_off,
1167 unsigned long *largest_bits)
1168{
1169 unsigned long index, end, i, area_off, area_bits;
1170again:
1171 index = find_next_zero_bit(map, size, start);
1172
1173 /* Align allocation */
1174 index = __ALIGN_MASK(index, align_mask);
1175 area_off = index;
1176
1177 end = index + nr;
1178 if (end > size)
1179 return end;
1180 i = find_next_bit(map, end, index);
1181 if (i < end) {
1182 area_bits = i - area_off;
1183 /* remember largest unused area with best alignment */
1184 if (area_bits > *largest_bits ||
1185 (area_bits == *largest_bits && *largest_off &&
1186 (!area_off || __ffs(area_off) > __ffs(*largest_off)))) {
1187 *largest_off = area_off;
1188 *largest_bits = area_bits;
1189 }
1190
1191 start = i + 1;
1192 goto again;
1193 }
1194 return index;
1195}
1196
1197/**
1198 * pcpu_alloc_area - allocates an area from a pcpu_chunk
1199 * @chunk: chunk of interest
1200 * @alloc_bits: size of request in allocation units
1201 * @align: alignment of area (max PAGE_SIZE)
1202 * @start: bit_off to start searching
1203 *
1204 * This function takes in a @start offset to begin searching to fit an
1205 * allocation of @alloc_bits with alignment @align. It needs to scan
1206 * the allocation map because if it fits within the block's contig hint,
1207 * @start will be block->first_free. This is an attempt to fill the
1208 * allocation prior to breaking the contig hint. The allocation and
1209 * boundary maps are updated accordingly if it confirms a valid
1210 * free area.
1211 *
1212 * RETURNS:
1213 * Allocated addr offset in @chunk on success.
1214 * -1 if no matching area is found.
1215 */
1216static int pcpu_alloc_area(struct pcpu_chunk *chunk, int alloc_bits,
1217 size_t align, int start)
1218{
1219 struct pcpu_block_md *chunk_md = &chunk->chunk_md;
1220 size_t align_mask = (align) ? (align - 1) : 0;
1221 unsigned long area_off = 0, area_bits = 0;
1222 int bit_off, end, oslot;
1223
1224 lockdep_assert_held(&pcpu_lock);
1225
1226 oslot = pcpu_chunk_slot(chunk);
1227
1228 /*
1229 * Search to find a fit.
1230 */
1231 end = min_t(int, start + alloc_bits + PCPU_BITMAP_BLOCK_BITS,
1232 pcpu_chunk_map_bits(chunk));
1233 bit_off = pcpu_find_zero_area(chunk->alloc_map, end, start, alloc_bits,
1234 align_mask, &area_off, &area_bits);
1235 if (bit_off >= end)
1236 return -1;
1237
1238 if (area_bits)
1239 pcpu_block_update_scan(chunk, area_off, area_bits);
1240
1241 /* update alloc map */
1242 bitmap_set(chunk->alloc_map, bit_off, alloc_bits);
1243
1244 /* update boundary map */
1245 set_bit(bit_off, chunk->bound_map);
1246 bitmap_clear(chunk->bound_map, bit_off + 1, alloc_bits - 1);
1247 set_bit(bit_off + alloc_bits, chunk->bound_map);
1248
1249 chunk->free_bytes -= alloc_bits * PCPU_MIN_ALLOC_SIZE;
1250
1251 /* update first free bit */
1252 if (bit_off == chunk_md->first_free)
1253 chunk_md->first_free = find_next_zero_bit(
1254 chunk->alloc_map,
1255 pcpu_chunk_map_bits(chunk),
1256 bit_off + alloc_bits);
1257
1258 pcpu_block_update_hint_alloc(chunk, bit_off, alloc_bits);
1259
1260 pcpu_chunk_relocate(chunk, oslot);
1261
1262 return bit_off * PCPU_MIN_ALLOC_SIZE;
1263}
1264
1265/**
1266 * pcpu_free_area - frees the corresponding offset
1267 * @chunk: chunk of interest
1268 * @off: addr offset into chunk
1269 *
1270 * This function determines the size of an allocation to free using
1271 * the boundary bitmap and clears the allocation map.
1272 *
1273 * RETURNS:
1274 * Number of freed bytes.
1275 */
1276static int pcpu_free_area(struct pcpu_chunk *chunk, int off)
1277{
1278 struct pcpu_block_md *chunk_md = &chunk->chunk_md;
1279 int bit_off, bits, end, oslot, freed;
1280
1281 lockdep_assert_held(&pcpu_lock);
1282 pcpu_stats_area_dealloc(chunk);
1283
1284 oslot = pcpu_chunk_slot(chunk);
1285
1286 bit_off = off / PCPU_MIN_ALLOC_SIZE;
1287
1288 /* find end index */
1289 end = find_next_bit(chunk->bound_map, pcpu_chunk_map_bits(chunk),
1290 bit_off + 1);
1291 bits = end - bit_off;
1292 bitmap_clear(chunk->alloc_map, bit_off, bits);
1293
1294 freed = bits * PCPU_MIN_ALLOC_SIZE;
1295
1296 /* update metadata */
1297 chunk->free_bytes += freed;
1298
1299 /* update first free bit */
1300 chunk_md->first_free = min(chunk_md->first_free, bit_off);
1301
1302 pcpu_block_update_hint_free(chunk, bit_off, bits);
1303
1304 pcpu_chunk_relocate(chunk, oslot);
1305
1306 return freed;
1307}
1308
1309static void pcpu_init_md_block(struct pcpu_block_md *block, int nr_bits)
1310{
1311 block->scan_hint = 0;
1312 block->contig_hint = nr_bits;
1313 block->left_free = nr_bits;
1314 block->right_free = nr_bits;
1315 block->first_free = 0;
1316 block->nr_bits = nr_bits;
1317}
1318
1319static void pcpu_init_md_blocks(struct pcpu_chunk *chunk)
1320{
1321 struct pcpu_block_md *md_block;
1322
1323 /* init the chunk's block */
1324 pcpu_init_md_block(&chunk->chunk_md, pcpu_chunk_map_bits(chunk));
1325
1326 for (md_block = chunk->md_blocks;
1327 md_block != chunk->md_blocks + pcpu_chunk_nr_blocks(chunk);
1328 md_block++)
1329 pcpu_init_md_block(md_block, PCPU_BITMAP_BLOCK_BITS);
1330}
1331
1332/**
1333 * pcpu_alloc_first_chunk - creates chunks that serve the first chunk
1334 * @tmp_addr: the start of the region served
1335 * @map_size: size of the region served
1336 *
1337 * This is responsible for creating the chunks that serve the first chunk. The
1338 * base_addr is page aligned down of @tmp_addr while the region end is page
1339 * aligned up. Offsets are kept track of to determine the region served. All
1340 * this is done to appease the bitmap allocator in avoiding partial blocks.
1341 *
1342 * RETURNS:
1343 * Chunk serving the region at @tmp_addr of @map_size.
1344 */
1345static struct pcpu_chunk * __init pcpu_alloc_first_chunk(unsigned long tmp_addr,
1346 int map_size)
1347{
1348 struct pcpu_chunk *chunk;
1349 unsigned long aligned_addr;
1350 int start_offset, offset_bits, region_size, region_bits;
1351 size_t alloc_size;
1352
1353 /* region calculations */
1354 aligned_addr = tmp_addr & PAGE_MASK;
1355
1356 start_offset = tmp_addr - aligned_addr;
1357 region_size = ALIGN(start_offset + map_size, PAGE_SIZE);
1358
1359 /* allocate chunk */
1360 alloc_size = struct_size(chunk, populated,
1361 BITS_TO_LONGS(region_size >> PAGE_SHIFT));
1362 chunk = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
1363 if (!chunk)
1364 panic("%s: Failed to allocate %zu bytes\n", __func__,
1365 alloc_size);
1366
1367 INIT_LIST_HEAD(&chunk->list);
1368
1369 chunk->base_addr = (void *)aligned_addr;
1370 chunk->start_offset = start_offset;
1371 chunk->end_offset = region_size - chunk->start_offset - map_size;
1372
1373 chunk->nr_pages = region_size >> PAGE_SHIFT;
1374 region_bits = pcpu_chunk_map_bits(chunk);
1375
1376 alloc_size = BITS_TO_LONGS(region_bits) * sizeof(chunk->alloc_map[0]);
1377 chunk->alloc_map = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
1378 if (!chunk->alloc_map)
1379 panic("%s: Failed to allocate %zu bytes\n", __func__,
1380 alloc_size);
1381
1382 alloc_size =
1383 BITS_TO_LONGS(region_bits + 1) * sizeof(chunk->bound_map[0]);
1384 chunk->bound_map = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
1385 if (!chunk->bound_map)
1386 panic("%s: Failed to allocate %zu bytes\n", __func__,
1387 alloc_size);
1388
1389 alloc_size = pcpu_chunk_nr_blocks(chunk) * sizeof(chunk->md_blocks[0]);
1390 chunk->md_blocks = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
1391 if (!chunk->md_blocks)
1392 panic("%s: Failed to allocate %zu bytes\n", __func__,
1393 alloc_size);
1394
1395#ifdef NEED_PCPUOBJ_EXT
1396 /* first chunk is free to use */
1397 chunk->obj_exts = NULL;
1398#endif
1399 pcpu_init_md_blocks(chunk);
1400
1401 /* manage populated page bitmap */
1402 chunk->immutable = true;
1403 bitmap_fill(chunk->populated, chunk->nr_pages);
1404 chunk->nr_populated = chunk->nr_pages;
1405 chunk->nr_empty_pop_pages = chunk->nr_pages;
1406
1407 chunk->free_bytes = map_size;
1408
1409 if (chunk->start_offset) {
1410 /* hide the beginning of the bitmap */
1411 offset_bits = chunk->start_offset / PCPU_MIN_ALLOC_SIZE;
1412 bitmap_set(chunk->alloc_map, 0, offset_bits);
1413 set_bit(0, chunk->bound_map);
1414 set_bit(offset_bits, chunk->bound_map);
1415
1416 chunk->chunk_md.first_free = offset_bits;
1417
1418 pcpu_block_update_hint_alloc(chunk, 0, offset_bits);
1419 }
1420
1421 if (chunk->end_offset) {
1422 /* hide the end of the bitmap */
1423 offset_bits = chunk->end_offset / PCPU_MIN_ALLOC_SIZE;
1424 bitmap_set(chunk->alloc_map,
1425 pcpu_chunk_map_bits(chunk) - offset_bits,
1426 offset_bits);
1427 set_bit((start_offset + map_size) / PCPU_MIN_ALLOC_SIZE,
1428 chunk->bound_map);
1429 set_bit(region_bits, chunk->bound_map);
1430
1431 pcpu_block_update_hint_alloc(chunk, pcpu_chunk_map_bits(chunk)
1432 - offset_bits, offset_bits);
1433 }
1434
1435 return chunk;
1436}
1437
1438static struct pcpu_chunk *pcpu_alloc_chunk(gfp_t gfp)
1439{
1440 struct pcpu_chunk *chunk;
1441 int region_bits;
1442
1443 chunk = pcpu_mem_zalloc(pcpu_chunk_struct_size, gfp);
1444 if (!chunk)
1445 return NULL;
1446
1447 INIT_LIST_HEAD(&chunk->list);
1448 chunk->nr_pages = pcpu_unit_pages;
1449 region_bits = pcpu_chunk_map_bits(chunk);
1450
1451 chunk->alloc_map = pcpu_mem_zalloc(BITS_TO_LONGS(region_bits) *
1452 sizeof(chunk->alloc_map[0]), gfp);
1453 if (!chunk->alloc_map)
1454 goto alloc_map_fail;
1455
1456 chunk->bound_map = pcpu_mem_zalloc(BITS_TO_LONGS(region_bits + 1) *
1457 sizeof(chunk->bound_map[0]), gfp);
1458 if (!chunk->bound_map)
1459 goto bound_map_fail;
1460
1461 chunk->md_blocks = pcpu_mem_zalloc(pcpu_chunk_nr_blocks(chunk) *
1462 sizeof(chunk->md_blocks[0]), gfp);
1463 if (!chunk->md_blocks)
1464 goto md_blocks_fail;
1465
1466#ifdef NEED_PCPUOBJ_EXT
1467 if (need_pcpuobj_ext()) {
1468 chunk->obj_exts =
1469 pcpu_mem_zalloc(pcpu_chunk_map_bits(chunk) *
1470 sizeof(struct pcpuobj_ext), gfp);
1471 if (!chunk->obj_exts)
1472 goto objcg_fail;
1473 }
1474#endif
1475
1476 pcpu_init_md_blocks(chunk);
1477
1478 /* init metadata */
1479 chunk->free_bytes = chunk->nr_pages * PAGE_SIZE;
1480
1481 return chunk;
1482
1483#ifdef NEED_PCPUOBJ_EXT
1484objcg_fail:
1485 pcpu_mem_free(chunk->md_blocks);
1486#endif
1487md_blocks_fail:
1488 pcpu_mem_free(chunk->bound_map);
1489bound_map_fail:
1490 pcpu_mem_free(chunk->alloc_map);
1491alloc_map_fail:
1492 pcpu_mem_free(chunk);
1493
1494 return NULL;
1495}
1496
1497static void pcpu_free_chunk(struct pcpu_chunk *chunk)
1498{
1499 if (!chunk)
1500 return;
1501#ifdef NEED_PCPUOBJ_EXT
1502 pcpu_mem_free(chunk->obj_exts);
1503#endif
1504 pcpu_mem_free(chunk->md_blocks);
1505 pcpu_mem_free(chunk->bound_map);
1506 pcpu_mem_free(chunk->alloc_map);
1507 pcpu_mem_free(chunk);
1508}
1509
1510/**
1511 * pcpu_chunk_populated - post-population bookkeeping
1512 * @chunk: pcpu_chunk which got populated
1513 * @page_start: the start page
1514 * @page_end: the end page
1515 *
1516 * Pages in [@page_start,@page_end) have been populated to @chunk. Update
1517 * the bookkeeping information accordingly. Must be called after each
1518 * successful population.
1519 */
1520static void pcpu_chunk_populated(struct pcpu_chunk *chunk, int page_start,
1521 int page_end)
1522{
1523 int nr = page_end - page_start;
1524
1525 lockdep_assert_held(&pcpu_lock);
1526
1527 bitmap_set(chunk->populated, page_start, nr);
1528 chunk->nr_populated += nr;
1529 pcpu_nr_populated += nr;
1530
1531 pcpu_update_empty_pages(chunk, nr);
1532}
1533
1534/**
1535 * pcpu_chunk_depopulated - post-depopulation bookkeeping
1536 * @chunk: pcpu_chunk which got depopulated
1537 * @page_start: the start page
1538 * @page_end: the end page
1539 *
1540 * Pages in [@page_start,@page_end) have been depopulated from @chunk.
1541 * Update the bookkeeping information accordingly. Must be called after
1542 * each successful depopulation.
1543 */
1544static void pcpu_chunk_depopulated(struct pcpu_chunk *chunk,
1545 int page_start, int page_end)
1546{
1547 int nr = page_end - page_start;
1548
1549 lockdep_assert_held(&pcpu_lock);
1550
1551 bitmap_clear(chunk->populated, page_start, nr);
1552 chunk->nr_populated -= nr;
1553 pcpu_nr_populated -= nr;
1554
1555 pcpu_update_empty_pages(chunk, -nr);
1556}
1557
1558/*
1559 * Chunk management implementation.
1560 *
1561 * To allow different implementations, chunk alloc/free and
1562 * [de]population are implemented in a separate file which is pulled
1563 * into this file and compiled together. The following functions
1564 * should be implemented.
1565 *
1566 * pcpu_populate_chunk - populate the specified range of a chunk
1567 * pcpu_depopulate_chunk - depopulate the specified range of a chunk
1568 * pcpu_post_unmap_tlb_flush - flush tlb for the specified range of a chunk
1569 * pcpu_create_chunk - create a new chunk
1570 * pcpu_destroy_chunk - destroy a chunk, always preceded by full depop
1571 * pcpu_addr_to_page - translate address to physical address
1572 * pcpu_verify_alloc_info - check alloc_info is acceptable during init
1573 */
1574static int pcpu_populate_chunk(struct pcpu_chunk *chunk,
1575 int page_start, int page_end, gfp_t gfp);
1576static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk,
1577 int page_start, int page_end);
1578static void pcpu_post_unmap_tlb_flush(struct pcpu_chunk *chunk,
1579 int page_start, int page_end);
1580static struct pcpu_chunk *pcpu_create_chunk(gfp_t gfp);
1581static void pcpu_destroy_chunk(struct pcpu_chunk *chunk);
1582static struct page *pcpu_addr_to_page(void *addr);
1583static int __init pcpu_verify_alloc_info(const struct pcpu_alloc_info *ai);
1584
1585#ifdef CONFIG_NEED_PER_CPU_KM
1586#include "percpu-km.c"
1587#else
1588#include "percpu-vm.c"
1589#endif
1590
1591/**
1592 * pcpu_chunk_addr_search - determine chunk containing specified address
1593 * @addr: address for which the chunk needs to be determined.
1594 *
1595 * This is an internal function that handles all but static allocations.
1596 * Static percpu address values should never be passed into the allocator.
1597 *
1598 * RETURNS:
1599 * The address of the found chunk.
1600 */
1601static struct pcpu_chunk *pcpu_chunk_addr_search(void *addr)
1602{
1603 /* is it in the dynamic region (first chunk)? */
1604 if (pcpu_addr_in_chunk(pcpu_first_chunk, addr))
1605 return pcpu_first_chunk;
1606
1607 /* is it in the reserved region? */
1608 if (pcpu_addr_in_chunk(pcpu_reserved_chunk, addr))
1609 return pcpu_reserved_chunk;
1610
1611 /*
1612 * The address is relative to unit0 which might be unused and
1613 * thus unmapped. Offset the address to the unit space of the
1614 * current processor before looking it up in the vmalloc
1615 * space. Note that any possible cpu id can be used here, so
1616 * there's no need to worry about preemption or cpu hotplug.
1617 */
1618 addr += pcpu_unit_offsets[raw_smp_processor_id()];
1619 return pcpu_get_page_chunk(pcpu_addr_to_page(addr));
1620}
1621
1622#ifdef CONFIG_MEMCG
1623static bool pcpu_memcg_pre_alloc_hook(size_t size, gfp_t gfp,
1624 struct obj_cgroup **objcgp)
1625{
1626 struct obj_cgroup *objcg;
1627
1628 if (!memcg_kmem_online() || !(gfp & __GFP_ACCOUNT))
1629 return true;
1630
1631 objcg = current_obj_cgroup();
1632 if (!objcg)
1633 return true;
1634
1635 if (obj_cgroup_charge(objcg, gfp, pcpu_obj_full_size(size)))
1636 return false;
1637
1638 *objcgp = objcg;
1639 return true;
1640}
1641
1642static void pcpu_memcg_post_alloc_hook(struct obj_cgroup *objcg,
1643 struct pcpu_chunk *chunk, int off,
1644 size_t size)
1645{
1646 if (!objcg)
1647 return;
1648
1649 if (likely(chunk && chunk->obj_exts)) {
1650 obj_cgroup_get(objcg);
1651 chunk->obj_exts[off >> PCPU_MIN_ALLOC_SHIFT].cgroup = objcg;
1652
1653 rcu_read_lock();
1654 mod_memcg_state(obj_cgroup_memcg(objcg), MEMCG_PERCPU_B,
1655 pcpu_obj_full_size(size));
1656 rcu_read_unlock();
1657 } else {
1658 obj_cgroup_uncharge(objcg, pcpu_obj_full_size(size));
1659 }
1660}
1661
1662static void pcpu_memcg_free_hook(struct pcpu_chunk *chunk, int off, size_t size)
1663{
1664 struct obj_cgroup *objcg;
1665
1666 if (unlikely(!chunk->obj_exts))
1667 return;
1668
1669 objcg = chunk->obj_exts[off >> PCPU_MIN_ALLOC_SHIFT].cgroup;
1670 if (!objcg)
1671 return;
1672 chunk->obj_exts[off >> PCPU_MIN_ALLOC_SHIFT].cgroup = NULL;
1673
1674 obj_cgroup_uncharge(objcg, pcpu_obj_full_size(size));
1675
1676 rcu_read_lock();
1677 mod_memcg_state(obj_cgroup_memcg(objcg), MEMCG_PERCPU_B,
1678 -pcpu_obj_full_size(size));
1679 rcu_read_unlock();
1680
1681 obj_cgroup_put(objcg);
1682}
1683
1684#else /* CONFIG_MEMCG */
1685static bool
1686pcpu_memcg_pre_alloc_hook(size_t size, gfp_t gfp, struct obj_cgroup **objcgp)
1687{
1688 return true;
1689}
1690
1691static void pcpu_memcg_post_alloc_hook(struct obj_cgroup *objcg,
1692 struct pcpu_chunk *chunk, int off,
1693 size_t size)
1694{
1695}
1696
1697static void pcpu_memcg_free_hook(struct pcpu_chunk *chunk, int off, size_t size)
1698{
1699}
1700#endif /* CONFIG_MEMCG */
1701
1702#ifdef CONFIG_MEM_ALLOC_PROFILING
1703static void pcpu_alloc_tag_alloc_hook(struct pcpu_chunk *chunk, int off,
1704 size_t size)
1705{
1706 if (mem_alloc_profiling_enabled() && likely(chunk->obj_exts)) {
1707 alloc_tag_add(&chunk->obj_exts[off >> PCPU_MIN_ALLOC_SHIFT].tag,
1708 current->alloc_tag, size);
1709 }
1710}
1711
1712static void pcpu_alloc_tag_free_hook(struct pcpu_chunk *chunk, int off, size_t size)
1713{
1714 if (mem_alloc_profiling_enabled() && likely(chunk->obj_exts))
1715 alloc_tag_sub(&chunk->obj_exts[off >> PCPU_MIN_ALLOC_SHIFT].tag, size);
1716}
1717#else
1718static void pcpu_alloc_tag_alloc_hook(struct pcpu_chunk *chunk, int off,
1719 size_t size)
1720{
1721}
1722
1723static void pcpu_alloc_tag_free_hook(struct pcpu_chunk *chunk, int off, size_t size)
1724{
1725}
1726#endif
1727
1728/**
1729 * pcpu_alloc - the percpu allocator
1730 * @size: size of area to allocate in bytes
1731 * @align: alignment of area (max PAGE_SIZE)
1732 * @reserved: allocate from the reserved chunk if available
1733 * @gfp: allocation flags
1734 *
1735 * Allocate percpu area of @size bytes aligned at @align. If @gfp doesn't
1736 * contain %GFP_KERNEL, the allocation is atomic. If @gfp has __GFP_NOWARN
1737 * then no warning will be triggered on invalid or failed allocation
1738 * requests.
1739 *
1740 * RETURNS:
1741 * Percpu pointer to the allocated area on success, NULL on failure.
1742 */
1743void __percpu *pcpu_alloc_noprof(size_t size, size_t align, bool reserved,
1744 gfp_t gfp)
1745{
1746 gfp_t pcpu_gfp;
1747 bool is_atomic;
1748 bool do_warn;
1749 struct obj_cgroup *objcg = NULL;
1750 static int warn_limit = 10;
1751 struct pcpu_chunk *chunk, *next;
1752 const char *err;
1753 int slot, off, cpu, ret;
1754 unsigned long flags;
1755 void __percpu *ptr;
1756 size_t bits, bit_align;
1757
1758 gfp = current_gfp_context(gfp);
1759 /* whitelisted flags that can be passed to the backing allocators */
1760 pcpu_gfp = gfp & (GFP_KERNEL | __GFP_NORETRY | __GFP_NOWARN);
1761 is_atomic = (gfp & GFP_KERNEL) != GFP_KERNEL;
1762 do_warn = !(gfp & __GFP_NOWARN);
1763
1764 /*
1765 * There is now a minimum allocation size of PCPU_MIN_ALLOC_SIZE,
1766 * therefore alignment must be a minimum of that many bytes.
1767 * An allocation may have internal fragmentation from rounding up
1768 * of up to PCPU_MIN_ALLOC_SIZE - 1 bytes.
1769 */
1770 if (unlikely(align < PCPU_MIN_ALLOC_SIZE))
1771 align = PCPU_MIN_ALLOC_SIZE;
1772
1773 size = ALIGN(size, PCPU_MIN_ALLOC_SIZE);
1774 bits = size >> PCPU_MIN_ALLOC_SHIFT;
1775 bit_align = align >> PCPU_MIN_ALLOC_SHIFT;
1776
1777 if (unlikely(!size || size > PCPU_MIN_UNIT_SIZE || align > PAGE_SIZE ||
1778 !is_power_of_2(align))) {
1779 WARN(do_warn, "illegal size (%zu) or align (%zu) for percpu allocation\n",
1780 size, align);
1781 return NULL;
1782 }
1783
1784 if (unlikely(!pcpu_memcg_pre_alloc_hook(size, gfp, &objcg)))
1785 return NULL;
1786
1787 if (!is_atomic) {
1788 /*
1789 * pcpu_balance_workfn() allocates memory under this mutex,
1790 * and it may wait for memory reclaim. Allow current task
1791 * to become OOM victim, in case of memory pressure.
1792 */
1793 if (gfp & __GFP_NOFAIL) {
1794 mutex_lock(&pcpu_alloc_mutex);
1795 } else if (mutex_lock_killable(&pcpu_alloc_mutex)) {
1796 pcpu_memcg_post_alloc_hook(objcg, NULL, 0, size);
1797 return NULL;
1798 }
1799 }
1800
1801 spin_lock_irqsave(&pcpu_lock, flags);
1802
1803 /* serve reserved allocations from the reserved chunk if available */
1804 if (reserved && pcpu_reserved_chunk) {
1805 chunk = pcpu_reserved_chunk;
1806
1807 off = pcpu_find_block_fit(chunk, bits, bit_align, is_atomic);
1808 if (off < 0) {
1809 err = "alloc from reserved chunk failed";
1810 goto fail_unlock;
1811 }
1812
1813 off = pcpu_alloc_area(chunk, bits, bit_align, off);
1814 if (off >= 0)
1815 goto area_found;
1816
1817 err = "alloc from reserved chunk failed";
1818 goto fail_unlock;
1819 }
1820
1821restart:
1822 /* search through normal chunks */
1823 for (slot = pcpu_size_to_slot(size); slot <= pcpu_free_slot; slot++) {
1824 list_for_each_entry_safe(chunk, next, &pcpu_chunk_lists[slot],
1825 list) {
1826 off = pcpu_find_block_fit(chunk, bits, bit_align,
1827 is_atomic);
1828 if (off < 0) {
1829 if (slot < PCPU_SLOT_FAIL_THRESHOLD)
1830 pcpu_chunk_move(chunk, 0);
1831 continue;
1832 }
1833
1834 off = pcpu_alloc_area(chunk, bits, bit_align, off);
1835 if (off >= 0) {
1836 pcpu_reintegrate_chunk(chunk);
1837 goto area_found;
1838 }
1839 }
1840 }
1841
1842 spin_unlock_irqrestore(&pcpu_lock, flags);
1843
1844 if (is_atomic) {
1845 err = "atomic alloc failed, no space left";
1846 goto fail;
1847 }
1848
1849 /* No space left. Create a new chunk. */
1850 if (list_empty(&pcpu_chunk_lists[pcpu_free_slot])) {
1851 chunk = pcpu_create_chunk(pcpu_gfp);
1852 if (!chunk) {
1853 err = "failed to allocate new chunk";
1854 goto fail;
1855 }
1856
1857 spin_lock_irqsave(&pcpu_lock, flags);
1858 pcpu_chunk_relocate(chunk, -1);
1859 } else {
1860 spin_lock_irqsave(&pcpu_lock, flags);
1861 }
1862
1863 goto restart;
1864
1865area_found:
1866 pcpu_stats_area_alloc(chunk, size);
1867
1868 if (pcpu_nr_empty_pop_pages < PCPU_EMPTY_POP_PAGES_LOW)
1869 pcpu_schedule_balance_work();
1870
1871 spin_unlock_irqrestore(&pcpu_lock, flags);
1872
1873 /* populate if not all pages are already there */
1874 if (!is_atomic) {
1875 unsigned int page_end, rs, re;
1876
1877 rs = PFN_DOWN(off);
1878 page_end = PFN_UP(off + size);
1879
1880 for_each_clear_bitrange_from(rs, re, chunk->populated, page_end) {
1881 WARN_ON(chunk->immutable);
1882
1883 ret = pcpu_populate_chunk(chunk, rs, re, pcpu_gfp);
1884
1885 spin_lock_irqsave(&pcpu_lock, flags);
1886 if (ret) {
1887 pcpu_free_area(chunk, off);
1888 err = "failed to populate";
1889 goto fail_unlock;
1890 }
1891 pcpu_chunk_populated(chunk, rs, re);
1892 spin_unlock_irqrestore(&pcpu_lock, flags);
1893 }
1894
1895 mutex_unlock(&pcpu_alloc_mutex);
1896 }
1897
1898 /* clear the areas and return address relative to base address */
1899 for_each_possible_cpu(cpu)
1900 memset((void *)pcpu_chunk_addr(chunk, cpu, 0) + off, 0, size);
1901
1902 ptr = __addr_to_pcpu_ptr(chunk->base_addr + off);
1903 kmemleak_alloc_percpu(ptr, size, gfp);
1904
1905 trace_percpu_alloc_percpu(_RET_IP_, reserved, is_atomic, size, align,
1906 chunk->base_addr, off, ptr,
1907 pcpu_obj_full_size(size), gfp);
1908
1909 pcpu_memcg_post_alloc_hook(objcg, chunk, off, size);
1910
1911 pcpu_alloc_tag_alloc_hook(chunk, off, size);
1912
1913 return ptr;
1914
1915fail_unlock:
1916 spin_unlock_irqrestore(&pcpu_lock, flags);
1917fail:
1918 trace_percpu_alloc_percpu_fail(reserved, is_atomic, size, align);
1919
1920 if (do_warn && warn_limit) {
1921 pr_warn("allocation failed, size=%zu align=%zu atomic=%d, %s\n",
1922 size, align, is_atomic, err);
1923 if (!is_atomic)
1924 dump_stack();
1925 if (!--warn_limit)
1926 pr_info("limit reached, disable warning\n");
1927 }
1928
1929 if (is_atomic) {
1930 /* see the flag handling in pcpu_balance_workfn() */
1931 pcpu_atomic_alloc_failed = true;
1932 pcpu_schedule_balance_work();
1933 } else {
1934 mutex_unlock(&pcpu_alloc_mutex);
1935 }
1936
1937 pcpu_memcg_post_alloc_hook(objcg, NULL, 0, size);
1938
1939 return NULL;
1940}
1941EXPORT_SYMBOL_GPL(pcpu_alloc_noprof);
1942
1943/**
1944 * pcpu_balance_free - manage the amount of free chunks
1945 * @empty_only: free chunks only if there are no populated pages
1946 *
1947 * If empty_only is %false, reclaim all fully free chunks regardless of the
1948 * number of populated pages. Otherwise, only reclaim chunks that have no
1949 * populated pages.
1950 *
1951 * CONTEXT:
1952 * pcpu_lock (can be dropped temporarily)
1953 */
1954static void pcpu_balance_free(bool empty_only)
1955{
1956 LIST_HEAD(to_free);
1957 struct list_head *free_head = &pcpu_chunk_lists[pcpu_free_slot];
1958 struct pcpu_chunk *chunk, *next;
1959
1960 lockdep_assert_held(&pcpu_lock);
1961
1962 /*
1963 * There's no reason to keep around multiple unused chunks and VM
1964 * areas can be scarce. Destroy all free chunks except for one.
1965 */
1966 list_for_each_entry_safe(chunk, next, free_head, list) {
1967 WARN_ON(chunk->immutable);
1968
1969 /* spare the first one */
1970 if (chunk == list_first_entry(free_head, struct pcpu_chunk, list))
1971 continue;
1972
1973 if (!empty_only || chunk->nr_empty_pop_pages == 0)
1974 list_move(&chunk->list, &to_free);
1975 }
1976
1977 if (list_empty(&to_free))
1978 return;
1979
1980 spin_unlock_irq(&pcpu_lock);
1981 list_for_each_entry_safe(chunk, next, &to_free, list) {
1982 unsigned int rs, re;
1983
1984 for_each_set_bitrange(rs, re, chunk->populated, chunk->nr_pages) {
1985 pcpu_depopulate_chunk(chunk, rs, re);
1986 spin_lock_irq(&pcpu_lock);
1987 pcpu_chunk_depopulated(chunk, rs, re);
1988 spin_unlock_irq(&pcpu_lock);
1989 }
1990 pcpu_destroy_chunk(chunk);
1991 cond_resched();
1992 }
1993 spin_lock_irq(&pcpu_lock);
1994}
1995
1996/**
1997 * pcpu_balance_populated - manage the amount of populated pages
1998 *
1999 * Maintain a certain amount of populated pages to satisfy atomic allocations.
2000 * It is possible that this is called when physical memory is scarce causing
2001 * OOM killer to be triggered. We should avoid doing so until an actual
2002 * allocation causes the failure as it is possible that requests can be
2003 * serviced from already backed regions.
2004 *
2005 * CONTEXT:
2006 * pcpu_lock (can be dropped temporarily)
2007 */
2008static void pcpu_balance_populated(void)
2009{
2010 /* gfp flags passed to underlying allocators */
2011 const gfp_t gfp = GFP_KERNEL | __GFP_NORETRY | __GFP_NOWARN;
2012 struct pcpu_chunk *chunk;
2013 int slot, nr_to_pop, ret;
2014
2015 lockdep_assert_held(&pcpu_lock);
2016
2017 /*
2018 * Ensure there are certain number of free populated pages for
2019 * atomic allocs. Fill up from the most packed so that atomic
2020 * allocs don't increase fragmentation. If atomic allocation
2021 * failed previously, always populate the maximum amount. This
2022 * should prevent atomic allocs larger than PAGE_SIZE from keeping
2023 * failing indefinitely; however, large atomic allocs are not
2024 * something we support properly and can be highly unreliable and
2025 * inefficient.
2026 */
2027retry_pop:
2028 if (pcpu_atomic_alloc_failed) {
2029 nr_to_pop = PCPU_EMPTY_POP_PAGES_HIGH;
2030 /* best effort anyway, don't worry about synchronization */
2031 pcpu_atomic_alloc_failed = false;
2032 } else {
2033 nr_to_pop = clamp(PCPU_EMPTY_POP_PAGES_HIGH -
2034 pcpu_nr_empty_pop_pages,
2035 0, PCPU_EMPTY_POP_PAGES_HIGH);
2036 }
2037
2038 for (slot = pcpu_size_to_slot(PAGE_SIZE); slot <= pcpu_free_slot; slot++) {
2039 unsigned int nr_unpop = 0, rs, re;
2040
2041 if (!nr_to_pop)
2042 break;
2043
2044 list_for_each_entry(chunk, &pcpu_chunk_lists[slot], list) {
2045 nr_unpop = chunk->nr_pages - chunk->nr_populated;
2046 if (nr_unpop)
2047 break;
2048 }
2049
2050 if (!nr_unpop)
2051 continue;
2052
2053 /* @chunk can't go away while pcpu_alloc_mutex is held */
2054 for_each_clear_bitrange(rs, re, chunk->populated, chunk->nr_pages) {
2055 int nr = min_t(int, re - rs, nr_to_pop);
2056
2057 spin_unlock_irq(&pcpu_lock);
2058 ret = pcpu_populate_chunk(chunk, rs, rs + nr, gfp);
2059 cond_resched();
2060 spin_lock_irq(&pcpu_lock);
2061 if (!ret) {
2062 nr_to_pop -= nr;
2063 pcpu_chunk_populated(chunk, rs, rs + nr);
2064 } else {
2065 nr_to_pop = 0;
2066 }
2067
2068 if (!nr_to_pop)
2069 break;
2070 }
2071 }
2072
2073 if (nr_to_pop) {
2074 /* ran out of chunks to populate, create a new one and retry */
2075 spin_unlock_irq(&pcpu_lock);
2076 chunk = pcpu_create_chunk(gfp);
2077 cond_resched();
2078 spin_lock_irq(&pcpu_lock);
2079 if (chunk) {
2080 pcpu_chunk_relocate(chunk, -1);
2081 goto retry_pop;
2082 }
2083 }
2084}
2085
2086/**
2087 * pcpu_reclaim_populated - scan over to_depopulate chunks and free empty pages
2088 *
2089 * Scan over chunks in the depopulate list and try to release unused populated
2090 * pages back to the system. Depopulated chunks are sidelined to prevent
2091 * repopulating these pages unless required. Fully free chunks are reintegrated
2092 * and freed accordingly (1 is kept around). If we drop below the empty
2093 * populated pages threshold, reintegrate the chunk if it has empty free pages.
2094 * Each chunk is scanned in the reverse order to keep populated pages close to
2095 * the beginning of the chunk.
2096 *
2097 * CONTEXT:
2098 * pcpu_lock (can be dropped temporarily)
2099 *
2100 */
2101static void pcpu_reclaim_populated(void)
2102{
2103 struct pcpu_chunk *chunk;
2104 struct pcpu_block_md *block;
2105 int freed_page_start, freed_page_end;
2106 int i, end;
2107 bool reintegrate;
2108
2109 lockdep_assert_held(&pcpu_lock);
2110
2111 /*
2112 * Once a chunk is isolated to the to_depopulate list, the chunk is no
2113 * longer discoverable to allocations whom may populate pages. The only
2114 * other accessor is the free path which only returns area back to the
2115 * allocator not touching the populated bitmap.
2116 */
2117 while ((chunk = list_first_entry_or_null(
2118 &pcpu_chunk_lists[pcpu_to_depopulate_slot],
2119 struct pcpu_chunk, list))) {
2120 WARN_ON(chunk->immutable);
2121
2122 /*
2123 * Scan chunk's pages in the reverse order to keep populated
2124 * pages close to the beginning of the chunk.
2125 */
2126 freed_page_start = chunk->nr_pages;
2127 freed_page_end = 0;
2128 reintegrate = false;
2129 for (i = chunk->nr_pages - 1, end = -1; i >= 0; i--) {
2130 /* no more work to do */
2131 if (chunk->nr_empty_pop_pages == 0)
2132 break;
2133
2134 /* reintegrate chunk to prevent atomic alloc failures */
2135 if (pcpu_nr_empty_pop_pages < PCPU_EMPTY_POP_PAGES_HIGH) {
2136 reintegrate = true;
2137 break;
2138 }
2139
2140 /*
2141 * If the page is empty and populated, start or
2142 * extend the (i, end) range. If i == 0, decrease
2143 * i and perform the depopulation to cover the last
2144 * (first) page in the chunk.
2145 */
2146 block = chunk->md_blocks + i;
2147 if (block->contig_hint == PCPU_BITMAP_BLOCK_BITS &&
2148 test_bit(i, chunk->populated)) {
2149 if (end == -1)
2150 end = i;
2151 if (i > 0)
2152 continue;
2153 i--;
2154 }
2155
2156 /* depopulate if there is an active range */
2157 if (end == -1)
2158 continue;
2159
2160 spin_unlock_irq(&pcpu_lock);
2161 pcpu_depopulate_chunk(chunk, i + 1, end + 1);
2162 cond_resched();
2163 spin_lock_irq(&pcpu_lock);
2164
2165 pcpu_chunk_depopulated(chunk, i + 1, end + 1);
2166 freed_page_start = min(freed_page_start, i + 1);
2167 freed_page_end = max(freed_page_end, end + 1);
2168
2169 /* reset the range and continue */
2170 end = -1;
2171 }
2172
2173 /* batch tlb flush per chunk to amortize cost */
2174 if (freed_page_start < freed_page_end) {
2175 spin_unlock_irq(&pcpu_lock);
2176 pcpu_post_unmap_tlb_flush(chunk,
2177 freed_page_start,
2178 freed_page_end);
2179 cond_resched();
2180 spin_lock_irq(&pcpu_lock);
2181 }
2182
2183 if (reintegrate || chunk->free_bytes == pcpu_unit_size)
2184 pcpu_reintegrate_chunk(chunk);
2185 else
2186 list_move_tail(&chunk->list,
2187 &pcpu_chunk_lists[pcpu_sidelined_slot]);
2188 }
2189}
2190
2191/**
2192 * pcpu_balance_workfn - manage the amount of free chunks and populated pages
2193 * @work: unused
2194 *
2195 * For each chunk type, manage the number of fully free chunks and the number of
2196 * populated pages. An important thing to consider is when pages are freed and
2197 * how they contribute to the global counts.
2198 */
2199static void pcpu_balance_workfn(struct work_struct *work)
2200{
2201 /*
2202 * pcpu_balance_free() is called twice because the first time we may
2203 * trim pages in the active pcpu_nr_empty_pop_pages which may cause us
2204 * to grow other chunks. This then gives pcpu_reclaim_populated() time
2205 * to move fully free chunks to the active list to be freed if
2206 * appropriate.
2207 */
2208 mutex_lock(&pcpu_alloc_mutex);
2209 spin_lock_irq(&pcpu_lock);
2210
2211 pcpu_balance_free(false);
2212 pcpu_reclaim_populated();
2213 pcpu_balance_populated();
2214 pcpu_balance_free(true);
2215
2216 spin_unlock_irq(&pcpu_lock);
2217 mutex_unlock(&pcpu_alloc_mutex);
2218}
2219
2220/**
2221 * free_percpu - free percpu area
2222 * @ptr: pointer to area to free
2223 *
2224 * Free percpu area @ptr.
2225 *
2226 * CONTEXT:
2227 * Can be called from atomic context.
2228 */
2229void free_percpu(void __percpu *ptr)
2230{
2231 void *addr;
2232 struct pcpu_chunk *chunk;
2233 unsigned long flags;
2234 int size, off;
2235 bool need_balance = false;
2236
2237 if (!ptr)
2238 return;
2239
2240 kmemleak_free_percpu(ptr);
2241
2242 addr = __pcpu_ptr_to_addr(ptr);
2243 chunk = pcpu_chunk_addr_search(addr);
2244 off = addr - chunk->base_addr;
2245
2246 spin_lock_irqsave(&pcpu_lock, flags);
2247 size = pcpu_free_area(chunk, off);
2248
2249 pcpu_alloc_tag_free_hook(chunk, off, size);
2250
2251 pcpu_memcg_free_hook(chunk, off, size);
2252
2253 /*
2254 * If there are more than one fully free chunks, wake up grim reaper.
2255 * If the chunk is isolated, it may be in the process of being
2256 * reclaimed. Let reclaim manage cleaning up of that chunk.
2257 */
2258 if (!chunk->isolated && chunk->free_bytes == pcpu_unit_size) {
2259 struct pcpu_chunk *pos;
2260
2261 list_for_each_entry(pos, &pcpu_chunk_lists[pcpu_free_slot], list)
2262 if (pos != chunk) {
2263 need_balance = true;
2264 break;
2265 }
2266 } else if (pcpu_should_reclaim_chunk(chunk)) {
2267 pcpu_isolate_chunk(chunk);
2268 need_balance = true;
2269 }
2270
2271 trace_percpu_free_percpu(chunk->base_addr, off, ptr);
2272
2273 spin_unlock_irqrestore(&pcpu_lock, flags);
2274
2275 if (need_balance)
2276 pcpu_schedule_balance_work();
2277}
2278EXPORT_SYMBOL_GPL(free_percpu);
2279
2280bool __is_kernel_percpu_address(unsigned long addr, unsigned long *can_addr)
2281{
2282#ifdef CONFIG_SMP
2283 const size_t static_size = __per_cpu_end - __per_cpu_start;
2284 void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr);
2285 unsigned int cpu;
2286
2287 for_each_possible_cpu(cpu) {
2288 void *start = per_cpu_ptr(base, cpu);
2289 void *va = (void *)addr;
2290
2291 if (va >= start && va < start + static_size) {
2292 if (can_addr) {
2293 *can_addr = (unsigned long) (va - start);
2294 *can_addr += (unsigned long)
2295 per_cpu_ptr(base, get_boot_cpu_id());
2296 }
2297 return true;
2298 }
2299 }
2300#endif
2301 /* on UP, can't distinguish from other static vars, always false */
2302 return false;
2303}
2304
2305/**
2306 * is_kernel_percpu_address - test whether address is from static percpu area
2307 * @addr: address to test
2308 *
2309 * Test whether @addr belongs to in-kernel static percpu area. Module
2310 * static percpu areas are not considered. For those, use
2311 * is_module_percpu_address().
2312 *
2313 * RETURNS:
2314 * %true if @addr is from in-kernel static percpu area, %false otherwise.
2315 */
2316bool is_kernel_percpu_address(unsigned long addr)
2317{
2318 return __is_kernel_percpu_address(addr, NULL);
2319}
2320
2321/**
2322 * per_cpu_ptr_to_phys - convert translated percpu address to physical address
2323 * @addr: the address to be converted to physical address
2324 *
2325 * Given @addr which is dereferenceable address obtained via one of
2326 * percpu access macros, this function translates it into its physical
2327 * address. The caller is responsible for ensuring @addr stays valid
2328 * until this function finishes.
2329 *
2330 * percpu allocator has special setup for the first chunk, which currently
2331 * supports either embedding in linear address space or vmalloc mapping,
2332 * and, from the second one, the backing allocator (currently either vm or
2333 * km) provides translation.
2334 *
2335 * The addr can be translated simply without checking if it falls into the
2336 * first chunk. But the current code reflects better how percpu allocator
2337 * actually works, and the verification can discover both bugs in percpu
2338 * allocator itself and per_cpu_ptr_to_phys() callers. So we keep current
2339 * code.
2340 *
2341 * RETURNS:
2342 * The physical address for @addr.
2343 */
2344phys_addr_t per_cpu_ptr_to_phys(void *addr)
2345{
2346 void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr);
2347 bool in_first_chunk = false;
2348 unsigned long first_low, first_high;
2349 unsigned int cpu;
2350
2351 /*
2352 * The following test on unit_low/high isn't strictly
2353 * necessary but will speed up lookups of addresses which
2354 * aren't in the first chunk.
2355 *
2356 * The address check is against full chunk sizes. pcpu_base_addr
2357 * points to the beginning of the first chunk including the
2358 * static region. Assumes good intent as the first chunk may
2359 * not be full (ie. < pcpu_unit_pages in size).
2360 */
2361 first_low = (unsigned long)pcpu_base_addr +
2362 pcpu_unit_page_offset(pcpu_low_unit_cpu, 0);
2363 first_high = (unsigned long)pcpu_base_addr +
2364 pcpu_unit_page_offset(pcpu_high_unit_cpu, pcpu_unit_pages);
2365 if ((unsigned long)addr >= first_low &&
2366 (unsigned long)addr < first_high) {
2367 for_each_possible_cpu(cpu) {
2368 void *start = per_cpu_ptr(base, cpu);
2369
2370 if (addr >= start && addr < start + pcpu_unit_size) {
2371 in_first_chunk = true;
2372 break;
2373 }
2374 }
2375 }
2376
2377 if (in_first_chunk) {
2378 if (!is_vmalloc_addr(addr))
2379 return __pa(addr);
2380 else
2381 return page_to_phys(vmalloc_to_page(addr)) +
2382 offset_in_page(addr);
2383 } else
2384 return page_to_phys(pcpu_addr_to_page(addr)) +
2385 offset_in_page(addr);
2386}
2387
2388/**
2389 * pcpu_alloc_alloc_info - allocate percpu allocation info
2390 * @nr_groups: the number of groups
2391 * @nr_units: the number of units
2392 *
2393 * Allocate ai which is large enough for @nr_groups groups containing
2394 * @nr_units units. The returned ai's groups[0].cpu_map points to the
2395 * cpu_map array which is long enough for @nr_units and filled with
2396 * NR_CPUS. It's the caller's responsibility to initialize cpu_map
2397 * pointer of other groups.
2398 *
2399 * RETURNS:
2400 * Pointer to the allocated pcpu_alloc_info on success, NULL on
2401 * failure.
2402 */
2403struct pcpu_alloc_info * __init pcpu_alloc_alloc_info(int nr_groups,
2404 int nr_units)
2405{
2406 struct pcpu_alloc_info *ai;
2407 size_t base_size, ai_size;
2408 void *ptr;
2409 int unit;
2410
2411 base_size = ALIGN(struct_size(ai, groups, nr_groups),
2412 __alignof__(ai->groups[0].cpu_map[0]));
2413 ai_size = base_size + nr_units * sizeof(ai->groups[0].cpu_map[0]);
2414
2415 ptr = memblock_alloc(PFN_ALIGN(ai_size), PAGE_SIZE);
2416 if (!ptr)
2417 return NULL;
2418 ai = ptr;
2419 ptr += base_size;
2420
2421 ai->groups[0].cpu_map = ptr;
2422
2423 for (unit = 0; unit < nr_units; unit++)
2424 ai->groups[0].cpu_map[unit] = NR_CPUS;
2425
2426 ai->nr_groups = nr_groups;
2427 ai->__ai_size = PFN_ALIGN(ai_size);
2428
2429 return ai;
2430}
2431
2432/**
2433 * pcpu_free_alloc_info - free percpu allocation info
2434 * @ai: pcpu_alloc_info to free
2435 *
2436 * Free @ai which was allocated by pcpu_alloc_alloc_info().
2437 */
2438void __init pcpu_free_alloc_info(struct pcpu_alloc_info *ai)
2439{
2440 memblock_free(ai, ai->__ai_size);
2441}
2442
2443/**
2444 * pcpu_dump_alloc_info - print out information about pcpu_alloc_info
2445 * @lvl: loglevel
2446 * @ai: allocation info to dump
2447 *
2448 * Print out information about @ai using loglevel @lvl.
2449 */
2450static void pcpu_dump_alloc_info(const char *lvl,
2451 const struct pcpu_alloc_info *ai)
2452{
2453 int group_width = 1, cpu_width = 1, width;
2454 char empty_str[] = "--------";
2455 int alloc = 0, alloc_end = 0;
2456 int group, v;
2457 int upa, apl; /* units per alloc, allocs per line */
2458
2459 v = ai->nr_groups;
2460 while (v /= 10)
2461 group_width++;
2462
2463 v = num_possible_cpus();
2464 while (v /= 10)
2465 cpu_width++;
2466 empty_str[min_t(int, cpu_width, sizeof(empty_str) - 1)] = '\0';
2467
2468 upa = ai->alloc_size / ai->unit_size;
2469 width = upa * (cpu_width + 1) + group_width + 3;
2470 apl = rounddown_pow_of_two(max(60 / width, 1));
2471
2472 printk("%spcpu-alloc: s%zu r%zu d%zu u%zu alloc=%zu*%zu",
2473 lvl, ai->static_size, ai->reserved_size, ai->dyn_size,
2474 ai->unit_size, ai->alloc_size / ai->atom_size, ai->atom_size);
2475
2476 for (group = 0; group < ai->nr_groups; group++) {
2477 const struct pcpu_group_info *gi = &ai->groups[group];
2478 int unit = 0, unit_end = 0;
2479
2480 BUG_ON(gi->nr_units % upa);
2481 for (alloc_end += gi->nr_units / upa;
2482 alloc < alloc_end; alloc++) {
2483 if (!(alloc % apl)) {
2484 pr_cont("\n");
2485 printk("%spcpu-alloc: ", lvl);
2486 }
2487 pr_cont("[%0*d] ", group_width, group);
2488
2489 for (unit_end += upa; unit < unit_end; unit++)
2490 if (gi->cpu_map[unit] != NR_CPUS)
2491 pr_cont("%0*d ",
2492 cpu_width, gi->cpu_map[unit]);
2493 else
2494 pr_cont("%s ", empty_str);
2495 }
2496 }
2497 pr_cont("\n");
2498}
2499
2500/**
2501 * pcpu_setup_first_chunk - initialize the first percpu chunk
2502 * @ai: pcpu_alloc_info describing how to percpu area is shaped
2503 * @base_addr: mapped address
2504 *
2505 * Initialize the first percpu chunk which contains the kernel static
2506 * percpu area. This function is to be called from arch percpu area
2507 * setup path.
2508 *
2509 * @ai contains all information necessary to initialize the first
2510 * chunk and prime the dynamic percpu allocator.
2511 *
2512 * @ai->static_size is the size of static percpu area.
2513 *
2514 * @ai->reserved_size, if non-zero, specifies the amount of bytes to
2515 * reserve after the static area in the first chunk. This reserves
2516 * the first chunk such that it's available only through reserved
2517 * percpu allocation. This is primarily used to serve module percpu
2518 * static areas on architectures where the addressing model has
2519 * limited offset range for symbol relocations to guarantee module
2520 * percpu symbols fall inside the relocatable range.
2521 *
2522 * @ai->dyn_size determines the number of bytes available for dynamic
2523 * allocation in the first chunk. The area between @ai->static_size +
2524 * @ai->reserved_size + @ai->dyn_size and @ai->unit_size is unused.
2525 *
2526 * @ai->unit_size specifies unit size and must be aligned to PAGE_SIZE
2527 * and equal to or larger than @ai->static_size + @ai->reserved_size +
2528 * @ai->dyn_size.
2529 *
2530 * @ai->atom_size is the allocation atom size and used as alignment
2531 * for vm areas.
2532 *
2533 * @ai->alloc_size is the allocation size and always multiple of
2534 * @ai->atom_size. This is larger than @ai->atom_size if
2535 * @ai->unit_size is larger than @ai->atom_size.
2536 *
2537 * @ai->nr_groups and @ai->groups describe virtual memory layout of
2538 * percpu areas. Units which should be colocated are put into the
2539 * same group. Dynamic VM areas will be allocated according to these
2540 * groupings. If @ai->nr_groups is zero, a single group containing
2541 * all units is assumed.
2542 *
2543 * The caller should have mapped the first chunk at @base_addr and
2544 * copied static data to each unit.
2545 *
2546 * The first chunk will always contain a static and a dynamic region.
2547 * However, the static region is not managed by any chunk. If the first
2548 * chunk also contains a reserved region, it is served by two chunks -
2549 * one for the reserved region and one for the dynamic region. They
2550 * share the same vm, but use offset regions in the area allocation map.
2551 * The chunk serving the dynamic region is circulated in the chunk slots
2552 * and available for dynamic allocation like any other chunk.
2553 */
2554void __init pcpu_setup_first_chunk(const struct pcpu_alloc_info *ai,
2555 void *base_addr)
2556{
2557 size_t size_sum = ai->static_size + ai->reserved_size + ai->dyn_size;
2558 size_t static_size, dyn_size;
2559 unsigned long *group_offsets;
2560 size_t *group_sizes;
2561 unsigned long *unit_off;
2562 unsigned int cpu;
2563 int *unit_map;
2564 int group, unit, i;
2565 unsigned long tmp_addr;
2566 size_t alloc_size;
2567
2568#define PCPU_SETUP_BUG_ON(cond) do { \
2569 if (unlikely(cond)) { \
2570 pr_emerg("failed to initialize, %s\n", #cond); \
2571 pr_emerg("cpu_possible_mask=%*pb\n", \
2572 cpumask_pr_args(cpu_possible_mask)); \
2573 pcpu_dump_alloc_info(KERN_EMERG, ai); \
2574 BUG(); \
2575 } \
2576} while (0)
2577
2578 /* sanity checks */
2579 PCPU_SETUP_BUG_ON(ai->nr_groups <= 0);
2580#ifdef CONFIG_SMP
2581 PCPU_SETUP_BUG_ON(!ai->static_size);
2582 PCPU_SETUP_BUG_ON(offset_in_page(__per_cpu_start));
2583#endif
2584 PCPU_SETUP_BUG_ON(!base_addr);
2585 PCPU_SETUP_BUG_ON(offset_in_page(base_addr));
2586 PCPU_SETUP_BUG_ON(ai->unit_size < size_sum);
2587 PCPU_SETUP_BUG_ON(offset_in_page(ai->unit_size));
2588 PCPU_SETUP_BUG_ON(ai->unit_size < PCPU_MIN_UNIT_SIZE);
2589 PCPU_SETUP_BUG_ON(!IS_ALIGNED(ai->unit_size, PCPU_BITMAP_BLOCK_SIZE));
2590 PCPU_SETUP_BUG_ON(ai->dyn_size < PERCPU_DYNAMIC_EARLY_SIZE);
2591 PCPU_SETUP_BUG_ON(!IS_ALIGNED(ai->reserved_size, PCPU_MIN_ALLOC_SIZE));
2592 PCPU_SETUP_BUG_ON(!(IS_ALIGNED(PCPU_BITMAP_BLOCK_SIZE, PAGE_SIZE) ||
2593 IS_ALIGNED(PAGE_SIZE, PCPU_BITMAP_BLOCK_SIZE)));
2594 PCPU_SETUP_BUG_ON(pcpu_verify_alloc_info(ai) < 0);
2595
2596 /* process group information and build config tables accordingly */
2597 alloc_size = ai->nr_groups * sizeof(group_offsets[0]);
2598 group_offsets = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
2599 if (!group_offsets)
2600 panic("%s: Failed to allocate %zu bytes\n", __func__,
2601 alloc_size);
2602
2603 alloc_size = ai->nr_groups * sizeof(group_sizes[0]);
2604 group_sizes = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
2605 if (!group_sizes)
2606 panic("%s: Failed to allocate %zu bytes\n", __func__,
2607 alloc_size);
2608
2609 alloc_size = nr_cpu_ids * sizeof(unit_map[0]);
2610 unit_map = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
2611 if (!unit_map)
2612 panic("%s: Failed to allocate %zu bytes\n", __func__,
2613 alloc_size);
2614
2615 alloc_size = nr_cpu_ids * sizeof(unit_off[0]);
2616 unit_off = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
2617 if (!unit_off)
2618 panic("%s: Failed to allocate %zu bytes\n", __func__,
2619 alloc_size);
2620
2621 for (cpu = 0; cpu < nr_cpu_ids; cpu++)
2622 unit_map[cpu] = UINT_MAX;
2623
2624 pcpu_low_unit_cpu = NR_CPUS;
2625 pcpu_high_unit_cpu = NR_CPUS;
2626
2627 for (group = 0, unit = 0; group < ai->nr_groups; group++, unit += i) {
2628 const struct pcpu_group_info *gi = &ai->groups[group];
2629
2630 group_offsets[group] = gi->base_offset;
2631 group_sizes[group] = gi->nr_units * ai->unit_size;
2632
2633 for (i = 0; i < gi->nr_units; i++) {
2634 cpu = gi->cpu_map[i];
2635 if (cpu == NR_CPUS)
2636 continue;
2637
2638 PCPU_SETUP_BUG_ON(cpu >= nr_cpu_ids);
2639 PCPU_SETUP_BUG_ON(!cpu_possible(cpu));
2640 PCPU_SETUP_BUG_ON(unit_map[cpu] != UINT_MAX);
2641
2642 unit_map[cpu] = unit + i;
2643 unit_off[cpu] = gi->base_offset + i * ai->unit_size;
2644
2645 /* determine low/high unit_cpu */
2646 if (pcpu_low_unit_cpu == NR_CPUS ||
2647 unit_off[cpu] < unit_off[pcpu_low_unit_cpu])
2648 pcpu_low_unit_cpu = cpu;
2649 if (pcpu_high_unit_cpu == NR_CPUS ||
2650 unit_off[cpu] > unit_off[pcpu_high_unit_cpu])
2651 pcpu_high_unit_cpu = cpu;
2652 }
2653 }
2654 pcpu_nr_units = unit;
2655
2656 for_each_possible_cpu(cpu)
2657 PCPU_SETUP_BUG_ON(unit_map[cpu] == UINT_MAX);
2658
2659 /* we're done parsing the input, undefine BUG macro and dump config */
2660#undef PCPU_SETUP_BUG_ON
2661 pcpu_dump_alloc_info(KERN_DEBUG, ai);
2662
2663 pcpu_nr_groups = ai->nr_groups;
2664 pcpu_group_offsets = group_offsets;
2665 pcpu_group_sizes = group_sizes;
2666 pcpu_unit_map = unit_map;
2667 pcpu_unit_offsets = unit_off;
2668
2669 /* determine basic parameters */
2670 pcpu_unit_pages = ai->unit_size >> PAGE_SHIFT;
2671 pcpu_unit_size = pcpu_unit_pages << PAGE_SHIFT;
2672 pcpu_atom_size = ai->atom_size;
2673 pcpu_chunk_struct_size = struct_size((struct pcpu_chunk *)0, populated,
2674 BITS_TO_LONGS(pcpu_unit_pages));
2675
2676 pcpu_stats_save_ai(ai);
2677
2678 /*
2679 * Allocate chunk slots. The slots after the active slots are:
2680 * sidelined_slot - isolated, depopulated chunks
2681 * free_slot - fully free chunks
2682 * to_depopulate_slot - isolated, chunks to depopulate
2683 */
2684 pcpu_sidelined_slot = __pcpu_size_to_slot(pcpu_unit_size) + 1;
2685 pcpu_free_slot = pcpu_sidelined_slot + 1;
2686 pcpu_to_depopulate_slot = pcpu_free_slot + 1;
2687 pcpu_nr_slots = pcpu_to_depopulate_slot + 1;
2688 pcpu_chunk_lists = memblock_alloc(pcpu_nr_slots *
2689 sizeof(pcpu_chunk_lists[0]),
2690 SMP_CACHE_BYTES);
2691 if (!pcpu_chunk_lists)
2692 panic("%s: Failed to allocate %zu bytes\n", __func__,
2693 pcpu_nr_slots * sizeof(pcpu_chunk_lists[0]));
2694
2695 for (i = 0; i < pcpu_nr_slots; i++)
2696 INIT_LIST_HEAD(&pcpu_chunk_lists[i]);
2697
2698 /*
2699 * The end of the static region needs to be aligned with the
2700 * minimum allocation size as this offsets the reserved and
2701 * dynamic region. The first chunk ends page aligned by
2702 * expanding the dynamic region, therefore the dynamic region
2703 * can be shrunk to compensate while still staying above the
2704 * configured sizes.
2705 */
2706 static_size = ALIGN(ai->static_size, PCPU_MIN_ALLOC_SIZE);
2707 dyn_size = ai->dyn_size - (static_size - ai->static_size);
2708
2709 /*
2710 * Initialize first chunk:
2711 * This chunk is broken up into 3 parts:
2712 * < static | [reserved] | dynamic >
2713 * - static - there is no backing chunk because these allocations can
2714 * never be freed.
2715 * - reserved (pcpu_reserved_chunk) - exists primarily to serve
2716 * allocations from module load.
2717 * - dynamic (pcpu_first_chunk) - serves the dynamic part of the first
2718 * chunk.
2719 */
2720 tmp_addr = (unsigned long)base_addr + static_size;
2721 if (ai->reserved_size)
2722 pcpu_reserved_chunk = pcpu_alloc_first_chunk(tmp_addr,
2723 ai->reserved_size);
2724 tmp_addr = (unsigned long)base_addr + static_size + ai->reserved_size;
2725 pcpu_first_chunk = pcpu_alloc_first_chunk(tmp_addr, dyn_size);
2726
2727 pcpu_nr_empty_pop_pages = pcpu_first_chunk->nr_empty_pop_pages;
2728 pcpu_chunk_relocate(pcpu_first_chunk, -1);
2729
2730 /* include all regions of the first chunk */
2731 pcpu_nr_populated += PFN_DOWN(size_sum);
2732
2733 pcpu_stats_chunk_alloc();
2734 trace_percpu_create_chunk(base_addr);
2735
2736 /* we're done */
2737 pcpu_base_addr = base_addr;
2738}
2739
2740#ifdef CONFIG_SMP
2741
2742const char * const pcpu_fc_names[PCPU_FC_NR] __initconst = {
2743 [PCPU_FC_AUTO] = "auto",
2744 [PCPU_FC_EMBED] = "embed",
2745 [PCPU_FC_PAGE] = "page",
2746};
2747
2748enum pcpu_fc pcpu_chosen_fc __initdata = PCPU_FC_AUTO;
2749
2750static int __init percpu_alloc_setup(char *str)
2751{
2752 if (!str)
2753 return -EINVAL;
2754
2755 if (0)
2756 /* nada */;
2757#ifdef CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK
2758 else if (!strcmp(str, "embed"))
2759 pcpu_chosen_fc = PCPU_FC_EMBED;
2760#endif
2761#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
2762 else if (!strcmp(str, "page"))
2763 pcpu_chosen_fc = PCPU_FC_PAGE;
2764#endif
2765 else
2766 pr_warn("unknown allocator %s specified\n", str);
2767
2768 return 0;
2769}
2770early_param("percpu_alloc", percpu_alloc_setup);
2771
2772/*
2773 * pcpu_embed_first_chunk() is used by the generic percpu setup.
2774 * Build it if needed by the arch config or the generic setup is going
2775 * to be used.
2776 */
2777#if defined(CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK) || \
2778 !defined(CONFIG_HAVE_SETUP_PER_CPU_AREA)
2779#define BUILD_EMBED_FIRST_CHUNK
2780#endif
2781
2782/* build pcpu_page_first_chunk() iff needed by the arch config */
2783#if defined(CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK)
2784#define BUILD_PAGE_FIRST_CHUNK
2785#endif
2786
2787/* pcpu_build_alloc_info() is used by both embed and page first chunk */
2788#if defined(BUILD_EMBED_FIRST_CHUNK) || defined(BUILD_PAGE_FIRST_CHUNK)
2789/**
2790 * pcpu_build_alloc_info - build alloc_info considering distances between CPUs
2791 * @reserved_size: the size of reserved percpu area in bytes
2792 * @dyn_size: minimum free size for dynamic allocation in bytes
2793 * @atom_size: allocation atom size
2794 * @cpu_distance_fn: callback to determine distance between cpus, optional
2795 *
2796 * This function determines grouping of units, their mappings to cpus
2797 * and other parameters considering needed percpu size, allocation
2798 * atom size and distances between CPUs.
2799 *
2800 * Groups are always multiples of atom size and CPUs which are of
2801 * LOCAL_DISTANCE both ways are grouped together and share space for
2802 * units in the same group. The returned configuration is guaranteed
2803 * to have CPUs on different nodes on different groups and >=75% usage
2804 * of allocated virtual address space.
2805 *
2806 * RETURNS:
2807 * On success, pointer to the new allocation_info is returned. On
2808 * failure, ERR_PTR value is returned.
2809 */
2810static struct pcpu_alloc_info * __init __flatten pcpu_build_alloc_info(
2811 size_t reserved_size, size_t dyn_size,
2812 size_t atom_size,
2813 pcpu_fc_cpu_distance_fn_t cpu_distance_fn)
2814{
2815 static int group_map[NR_CPUS] __initdata;
2816 static int group_cnt[NR_CPUS] __initdata;
2817 static struct cpumask mask __initdata;
2818 const size_t static_size = __per_cpu_end - __per_cpu_start;
2819 int nr_groups = 1, nr_units = 0;
2820 size_t size_sum, min_unit_size, alloc_size;
2821 int upa, max_upa, best_upa; /* units_per_alloc */
2822 int last_allocs, group, unit;
2823 unsigned int cpu, tcpu;
2824 struct pcpu_alloc_info *ai;
2825 unsigned int *cpu_map;
2826
2827 /* this function may be called multiple times */
2828 memset(group_map, 0, sizeof(group_map));
2829 memset(group_cnt, 0, sizeof(group_cnt));
2830 cpumask_clear(&mask);
2831
2832 /* calculate size_sum and ensure dyn_size is enough for early alloc */
2833 size_sum = PFN_ALIGN(static_size + reserved_size +
2834 max_t(size_t, dyn_size, PERCPU_DYNAMIC_EARLY_SIZE));
2835 dyn_size = size_sum - static_size - reserved_size;
2836
2837 /*
2838 * Determine min_unit_size, alloc_size and max_upa such that
2839 * alloc_size is multiple of atom_size and is the smallest
2840 * which can accommodate 4k aligned segments which are equal to
2841 * or larger than min_unit_size.
2842 */
2843 min_unit_size = max_t(size_t, size_sum, PCPU_MIN_UNIT_SIZE);
2844
2845 /* determine the maximum # of units that can fit in an allocation */
2846 alloc_size = roundup(min_unit_size, atom_size);
2847 upa = alloc_size / min_unit_size;
2848 while (alloc_size % upa || (offset_in_page(alloc_size / upa)))
2849 upa--;
2850 max_upa = upa;
2851
2852 cpumask_copy(&mask, cpu_possible_mask);
2853
2854 /* group cpus according to their proximity */
2855 for (group = 0; !cpumask_empty(&mask); group++) {
2856 /* pop the group's first cpu */
2857 cpu = cpumask_first(&mask);
2858 group_map[cpu] = group;
2859 group_cnt[group]++;
2860 cpumask_clear_cpu(cpu, &mask);
2861
2862 for_each_cpu(tcpu, &mask) {
2863 if (!cpu_distance_fn ||
2864 (cpu_distance_fn(cpu, tcpu) == LOCAL_DISTANCE &&
2865 cpu_distance_fn(tcpu, cpu) == LOCAL_DISTANCE)) {
2866 group_map[tcpu] = group;
2867 group_cnt[group]++;
2868 cpumask_clear_cpu(tcpu, &mask);
2869 }
2870 }
2871 }
2872 nr_groups = group;
2873
2874 /*
2875 * Wasted space is caused by a ratio imbalance of upa to group_cnt.
2876 * Expand the unit_size until we use >= 75% of the units allocated.
2877 * Related to atom_size, which could be much larger than the unit_size.
2878 */
2879 last_allocs = INT_MAX;
2880 best_upa = 0;
2881 for (upa = max_upa; upa; upa--) {
2882 int allocs = 0, wasted = 0;
2883
2884 if (alloc_size % upa || (offset_in_page(alloc_size / upa)))
2885 continue;
2886
2887 for (group = 0; group < nr_groups; group++) {
2888 int this_allocs = DIV_ROUND_UP(group_cnt[group], upa);
2889 allocs += this_allocs;
2890 wasted += this_allocs * upa - group_cnt[group];
2891 }
2892
2893 /*
2894 * Don't accept if wastage is over 1/3. The
2895 * greater-than comparison ensures upa==1 always
2896 * passes the following check.
2897 */
2898 if (wasted > num_possible_cpus() / 3)
2899 continue;
2900
2901 /* and then don't consume more memory */
2902 if (allocs > last_allocs)
2903 break;
2904 last_allocs = allocs;
2905 best_upa = upa;
2906 }
2907 BUG_ON(!best_upa);
2908 upa = best_upa;
2909
2910 /* allocate and fill alloc_info */
2911 for (group = 0; group < nr_groups; group++)
2912 nr_units += roundup(group_cnt[group], upa);
2913
2914 ai = pcpu_alloc_alloc_info(nr_groups, nr_units);
2915 if (!ai)
2916 return ERR_PTR(-ENOMEM);
2917 cpu_map = ai->groups[0].cpu_map;
2918
2919 for (group = 0; group < nr_groups; group++) {
2920 ai->groups[group].cpu_map = cpu_map;
2921 cpu_map += roundup(group_cnt[group], upa);
2922 }
2923
2924 ai->static_size = static_size;
2925 ai->reserved_size = reserved_size;
2926 ai->dyn_size = dyn_size;
2927 ai->unit_size = alloc_size / upa;
2928 ai->atom_size = atom_size;
2929 ai->alloc_size = alloc_size;
2930
2931 for (group = 0, unit = 0; group < nr_groups; group++) {
2932 struct pcpu_group_info *gi = &ai->groups[group];
2933
2934 /*
2935 * Initialize base_offset as if all groups are located
2936 * back-to-back. The caller should update this to
2937 * reflect actual allocation.
2938 */
2939 gi->base_offset = unit * ai->unit_size;
2940
2941 for_each_possible_cpu(cpu)
2942 if (group_map[cpu] == group)
2943 gi->cpu_map[gi->nr_units++] = cpu;
2944 gi->nr_units = roundup(gi->nr_units, upa);
2945 unit += gi->nr_units;
2946 }
2947 BUG_ON(unit != nr_units);
2948
2949 return ai;
2950}
2951
2952static void * __init pcpu_fc_alloc(unsigned int cpu, size_t size, size_t align,
2953 pcpu_fc_cpu_to_node_fn_t cpu_to_nd_fn)
2954{
2955 const unsigned long goal = __pa(MAX_DMA_ADDRESS);
2956#ifdef CONFIG_NUMA
2957 int node = NUMA_NO_NODE;
2958 void *ptr;
2959
2960 if (cpu_to_nd_fn)
2961 node = cpu_to_nd_fn(cpu);
2962
2963 if (node == NUMA_NO_NODE || !node_online(node) || !NODE_DATA(node)) {
2964 ptr = memblock_alloc_from(size, align, goal);
2965 pr_info("cpu %d has no node %d or node-local memory\n",
2966 cpu, node);
2967 pr_debug("per cpu data for cpu%d %zu bytes at 0x%llx\n",
2968 cpu, size, (u64)__pa(ptr));
2969 } else {
2970 ptr = memblock_alloc_try_nid(size, align, goal,
2971 MEMBLOCK_ALLOC_ACCESSIBLE,
2972 node);
2973
2974 pr_debug("per cpu data for cpu%d %zu bytes on node%d at 0x%llx\n",
2975 cpu, size, node, (u64)__pa(ptr));
2976 }
2977 return ptr;
2978#else
2979 return memblock_alloc_from(size, align, goal);
2980#endif
2981}
2982
2983static void __init pcpu_fc_free(void *ptr, size_t size)
2984{
2985 memblock_free(ptr, size);
2986}
2987#endif /* BUILD_EMBED_FIRST_CHUNK || BUILD_PAGE_FIRST_CHUNK */
2988
2989#if defined(BUILD_EMBED_FIRST_CHUNK)
2990/**
2991 * pcpu_embed_first_chunk - embed the first percpu chunk into bootmem
2992 * @reserved_size: the size of reserved percpu area in bytes
2993 * @dyn_size: minimum free size for dynamic allocation in bytes
2994 * @atom_size: allocation atom size
2995 * @cpu_distance_fn: callback to determine distance between cpus, optional
2996 * @cpu_to_nd_fn: callback to convert cpu to it's node, optional
2997 *
2998 * This is a helper to ease setting up embedded first percpu chunk and
2999 * can be called where pcpu_setup_first_chunk() is expected.
3000 *
3001 * If this function is used to setup the first chunk, it is allocated
3002 * by calling pcpu_fc_alloc and used as-is without being mapped into
3003 * vmalloc area. Allocations are always whole multiples of @atom_size
3004 * aligned to @atom_size.
3005 *
3006 * This enables the first chunk to piggy back on the linear physical
3007 * mapping which often uses larger page size. Please note that this
3008 * can result in very sparse cpu->unit mapping on NUMA machines thus
3009 * requiring large vmalloc address space. Don't use this allocator if
3010 * vmalloc space is not orders of magnitude larger than distances
3011 * between node memory addresses (ie. 32bit NUMA machines).
3012 *
3013 * @dyn_size specifies the minimum dynamic area size.
3014 *
3015 * If the needed size is smaller than the minimum or specified unit
3016 * size, the leftover is returned using pcpu_fc_free.
3017 *
3018 * RETURNS:
3019 * 0 on success, -errno on failure.
3020 */
3021int __init pcpu_embed_first_chunk(size_t reserved_size, size_t dyn_size,
3022 size_t atom_size,
3023 pcpu_fc_cpu_distance_fn_t cpu_distance_fn,
3024 pcpu_fc_cpu_to_node_fn_t cpu_to_nd_fn)
3025{
3026 void *base = (void *)ULONG_MAX;
3027 void **areas = NULL;
3028 struct pcpu_alloc_info *ai;
3029 size_t size_sum, areas_size;
3030 unsigned long max_distance;
3031 int group, i, highest_group, rc = 0;
3032
3033 ai = pcpu_build_alloc_info(reserved_size, dyn_size, atom_size,
3034 cpu_distance_fn);
3035 if (IS_ERR(ai))
3036 return PTR_ERR(ai);
3037
3038 size_sum = ai->static_size + ai->reserved_size + ai->dyn_size;
3039 areas_size = PFN_ALIGN(ai->nr_groups * sizeof(void *));
3040
3041 areas = memblock_alloc(areas_size, SMP_CACHE_BYTES);
3042 if (!areas) {
3043 rc = -ENOMEM;
3044 goto out_free;
3045 }
3046
3047 /* allocate, copy and determine base address & max_distance */
3048 highest_group = 0;
3049 for (group = 0; group < ai->nr_groups; group++) {
3050 struct pcpu_group_info *gi = &ai->groups[group];
3051 unsigned int cpu = NR_CPUS;
3052 void *ptr;
3053
3054 for (i = 0; i < gi->nr_units && cpu == NR_CPUS; i++)
3055 cpu = gi->cpu_map[i];
3056 BUG_ON(cpu == NR_CPUS);
3057
3058 /* allocate space for the whole group */
3059 ptr = pcpu_fc_alloc(cpu, gi->nr_units * ai->unit_size, atom_size, cpu_to_nd_fn);
3060 if (!ptr) {
3061 rc = -ENOMEM;
3062 goto out_free_areas;
3063 }
3064 /* kmemleak tracks the percpu allocations separately */
3065 kmemleak_ignore_phys(__pa(ptr));
3066 areas[group] = ptr;
3067
3068 base = min(ptr, base);
3069 if (ptr > areas[highest_group])
3070 highest_group = group;
3071 }
3072 max_distance = areas[highest_group] - base;
3073 max_distance += ai->unit_size * ai->groups[highest_group].nr_units;
3074
3075 /* warn if maximum distance is further than 75% of vmalloc space */
3076 if (max_distance > VMALLOC_TOTAL * 3 / 4) {
3077 pr_warn("max_distance=0x%lx too large for vmalloc space 0x%lx\n",
3078 max_distance, VMALLOC_TOTAL);
3079#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
3080 /* and fail if we have fallback */
3081 rc = -EINVAL;
3082 goto out_free_areas;
3083#endif
3084 }
3085
3086 /*
3087 * Copy data and free unused parts. This should happen after all
3088 * allocations are complete; otherwise, we may end up with
3089 * overlapping groups.
3090 */
3091 for (group = 0; group < ai->nr_groups; group++) {
3092 struct pcpu_group_info *gi = &ai->groups[group];
3093 void *ptr = areas[group];
3094
3095 for (i = 0; i < gi->nr_units; i++, ptr += ai->unit_size) {
3096 if (gi->cpu_map[i] == NR_CPUS) {
3097 /* unused unit, free whole */
3098 pcpu_fc_free(ptr, ai->unit_size);
3099 continue;
3100 }
3101 /* copy and return the unused part */
3102 memcpy(ptr, __per_cpu_load, ai->static_size);
3103 pcpu_fc_free(ptr + size_sum, ai->unit_size - size_sum);
3104 }
3105 }
3106
3107 /* base address is now known, determine group base offsets */
3108 for (group = 0; group < ai->nr_groups; group++) {
3109 ai->groups[group].base_offset = areas[group] - base;
3110 }
3111
3112 pr_info("Embedded %zu pages/cpu s%zu r%zu d%zu u%zu\n",
3113 PFN_DOWN(size_sum), ai->static_size, ai->reserved_size,
3114 ai->dyn_size, ai->unit_size);
3115
3116 pcpu_setup_first_chunk(ai, base);
3117 goto out_free;
3118
3119out_free_areas:
3120 for (group = 0; group < ai->nr_groups; group++)
3121 if (areas[group])
3122 pcpu_fc_free(areas[group],
3123 ai->groups[group].nr_units * ai->unit_size);
3124out_free:
3125 pcpu_free_alloc_info(ai);
3126 if (areas)
3127 memblock_free(areas, areas_size);
3128 return rc;
3129}
3130#endif /* BUILD_EMBED_FIRST_CHUNK */
3131
3132#ifdef BUILD_PAGE_FIRST_CHUNK
3133#include <asm/pgalloc.h>
3134
3135#ifndef P4D_TABLE_SIZE
3136#define P4D_TABLE_SIZE PAGE_SIZE
3137#endif
3138
3139#ifndef PUD_TABLE_SIZE
3140#define PUD_TABLE_SIZE PAGE_SIZE
3141#endif
3142
3143#ifndef PMD_TABLE_SIZE
3144#define PMD_TABLE_SIZE PAGE_SIZE
3145#endif
3146
3147#ifndef PTE_TABLE_SIZE
3148#define PTE_TABLE_SIZE PAGE_SIZE
3149#endif
3150void __init __weak pcpu_populate_pte(unsigned long addr)
3151{
3152 pgd_t *pgd = pgd_offset_k(addr);
3153 p4d_t *p4d;
3154 pud_t *pud;
3155 pmd_t *pmd;
3156
3157 if (pgd_none(*pgd)) {
3158 p4d = memblock_alloc(P4D_TABLE_SIZE, P4D_TABLE_SIZE);
3159 if (!p4d)
3160 goto err_alloc;
3161 pgd_populate(&init_mm, pgd, p4d);
3162 }
3163
3164 p4d = p4d_offset(pgd, addr);
3165 if (p4d_none(*p4d)) {
3166 pud = memblock_alloc(PUD_TABLE_SIZE, PUD_TABLE_SIZE);
3167 if (!pud)
3168 goto err_alloc;
3169 p4d_populate(&init_mm, p4d, pud);
3170 }
3171
3172 pud = pud_offset(p4d, addr);
3173 if (pud_none(*pud)) {
3174 pmd = memblock_alloc(PMD_TABLE_SIZE, PMD_TABLE_SIZE);
3175 if (!pmd)
3176 goto err_alloc;
3177 pud_populate(&init_mm, pud, pmd);
3178 }
3179
3180 pmd = pmd_offset(pud, addr);
3181 if (!pmd_present(*pmd)) {
3182 pte_t *new;
3183
3184 new = memblock_alloc(PTE_TABLE_SIZE, PTE_TABLE_SIZE);
3185 if (!new)
3186 goto err_alloc;
3187 pmd_populate_kernel(&init_mm, pmd, new);
3188 }
3189
3190 return;
3191
3192err_alloc:
3193 panic("%s: Failed to allocate memory\n", __func__);
3194}
3195
3196/**
3197 * pcpu_page_first_chunk - map the first chunk using PAGE_SIZE pages
3198 * @reserved_size: the size of reserved percpu area in bytes
3199 * @cpu_to_nd_fn: callback to convert cpu to it's node, optional
3200 *
3201 * This is a helper to ease setting up page-remapped first percpu
3202 * chunk and can be called where pcpu_setup_first_chunk() is expected.
3203 *
3204 * This is the basic allocator. Static percpu area is allocated
3205 * page-by-page into vmalloc area.
3206 *
3207 * RETURNS:
3208 * 0 on success, -errno on failure.
3209 */
3210int __init pcpu_page_first_chunk(size_t reserved_size, pcpu_fc_cpu_to_node_fn_t cpu_to_nd_fn)
3211{
3212 static struct vm_struct vm;
3213 struct pcpu_alloc_info *ai;
3214 char psize_str[16];
3215 int unit_pages;
3216 size_t pages_size;
3217 struct page **pages;
3218 int unit, i, j, rc = 0;
3219 int upa;
3220 int nr_g0_units;
3221
3222 snprintf(psize_str, sizeof(psize_str), "%luK", PAGE_SIZE >> 10);
3223
3224 ai = pcpu_build_alloc_info(reserved_size, 0, PAGE_SIZE, NULL);
3225 if (IS_ERR(ai))
3226 return PTR_ERR(ai);
3227 BUG_ON(ai->nr_groups != 1);
3228 upa = ai->alloc_size/ai->unit_size;
3229 nr_g0_units = roundup(num_possible_cpus(), upa);
3230 if (WARN_ON(ai->groups[0].nr_units != nr_g0_units)) {
3231 pcpu_free_alloc_info(ai);
3232 return -EINVAL;
3233 }
3234
3235 unit_pages = ai->unit_size >> PAGE_SHIFT;
3236
3237 /* unaligned allocations can't be freed, round up to page size */
3238 pages_size = PFN_ALIGN(unit_pages * num_possible_cpus() *
3239 sizeof(pages[0]));
3240 pages = memblock_alloc(pages_size, SMP_CACHE_BYTES);
3241 if (!pages)
3242 panic("%s: Failed to allocate %zu bytes\n", __func__,
3243 pages_size);
3244
3245 /* allocate pages */
3246 j = 0;
3247 for (unit = 0; unit < num_possible_cpus(); unit++) {
3248 unsigned int cpu = ai->groups[0].cpu_map[unit];
3249 for (i = 0; i < unit_pages; i++) {
3250 void *ptr;
3251
3252 ptr = pcpu_fc_alloc(cpu, PAGE_SIZE, PAGE_SIZE, cpu_to_nd_fn);
3253 if (!ptr) {
3254 pr_warn("failed to allocate %s page for cpu%u\n",
3255 psize_str, cpu);
3256 goto enomem;
3257 }
3258 /* kmemleak tracks the percpu allocations separately */
3259 kmemleak_ignore_phys(__pa(ptr));
3260 pages[j++] = virt_to_page(ptr);
3261 }
3262 }
3263
3264 /* allocate vm area, map the pages and copy static data */
3265 vm.flags = VM_ALLOC;
3266 vm.size = num_possible_cpus() * ai->unit_size;
3267 vm_area_register_early(&vm, PAGE_SIZE);
3268
3269 for (unit = 0; unit < num_possible_cpus(); unit++) {
3270 unsigned long unit_addr =
3271 (unsigned long)vm.addr + unit * ai->unit_size;
3272
3273 for (i = 0; i < unit_pages; i++)
3274 pcpu_populate_pte(unit_addr + (i << PAGE_SHIFT));
3275
3276 /* pte already populated, the following shouldn't fail */
3277 rc = __pcpu_map_pages(unit_addr, &pages[unit * unit_pages],
3278 unit_pages);
3279 if (rc < 0)
3280 panic("failed to map percpu area, err=%d\n", rc);
3281
3282 flush_cache_vmap_early(unit_addr, unit_addr + ai->unit_size);
3283
3284 /* copy static data */
3285 memcpy((void *)unit_addr, __per_cpu_load, ai->static_size);
3286 }
3287
3288 /* we're ready, commit */
3289 pr_info("%d %s pages/cpu s%zu r%zu d%zu\n",
3290 unit_pages, psize_str, ai->static_size,
3291 ai->reserved_size, ai->dyn_size);
3292
3293 pcpu_setup_first_chunk(ai, vm.addr);
3294 goto out_free_ar;
3295
3296enomem:
3297 while (--j >= 0)
3298 pcpu_fc_free(page_address(pages[j]), PAGE_SIZE);
3299 rc = -ENOMEM;
3300out_free_ar:
3301 memblock_free(pages, pages_size);
3302 pcpu_free_alloc_info(ai);
3303 return rc;
3304}
3305#endif /* BUILD_PAGE_FIRST_CHUNK */
3306
3307#ifndef CONFIG_HAVE_SETUP_PER_CPU_AREA
3308/*
3309 * Generic SMP percpu area setup.
3310 *
3311 * The embedding helper is used because its behavior closely resembles
3312 * the original non-dynamic generic percpu area setup. This is
3313 * important because many archs have addressing restrictions and might
3314 * fail if the percpu area is located far away from the previous
3315 * location. As an added bonus, in non-NUMA cases, embedding is
3316 * generally a good idea TLB-wise because percpu area can piggy back
3317 * on the physical linear memory mapping which uses large page
3318 * mappings on applicable archs.
3319 */
3320unsigned long __per_cpu_offset[NR_CPUS] __read_mostly;
3321EXPORT_SYMBOL(__per_cpu_offset);
3322
3323void __init setup_per_cpu_areas(void)
3324{
3325 unsigned long delta;
3326 unsigned int cpu;
3327 int rc;
3328
3329 /*
3330 * Always reserve area for module percpu variables. That's
3331 * what the legacy allocator did.
3332 */
3333 rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE, PERCPU_DYNAMIC_RESERVE,
3334 PAGE_SIZE, NULL, NULL);
3335 if (rc < 0)
3336 panic("Failed to initialize percpu areas.");
3337
3338 delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
3339 for_each_possible_cpu(cpu)
3340 __per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu];
3341}
3342#endif /* CONFIG_HAVE_SETUP_PER_CPU_AREA */
3343
3344#else /* CONFIG_SMP */
3345
3346/*
3347 * UP percpu area setup.
3348 *
3349 * UP always uses km-based percpu allocator with identity mapping.
3350 * Static percpu variables are indistinguishable from the usual static
3351 * variables and don't require any special preparation.
3352 */
3353void __init setup_per_cpu_areas(void)
3354{
3355 const size_t unit_size =
3356 roundup_pow_of_two(max_t(size_t, PCPU_MIN_UNIT_SIZE,
3357 PERCPU_DYNAMIC_RESERVE));
3358 struct pcpu_alloc_info *ai;
3359 void *fc;
3360
3361 ai = pcpu_alloc_alloc_info(1, 1);
3362 fc = memblock_alloc_from(unit_size, PAGE_SIZE, __pa(MAX_DMA_ADDRESS));
3363 if (!ai || !fc)
3364 panic("Failed to allocate memory for percpu areas.");
3365 /* kmemleak tracks the percpu allocations separately */
3366 kmemleak_ignore_phys(__pa(fc));
3367
3368 ai->dyn_size = unit_size;
3369 ai->unit_size = unit_size;
3370 ai->atom_size = unit_size;
3371 ai->alloc_size = unit_size;
3372 ai->groups[0].nr_units = 1;
3373 ai->groups[0].cpu_map[0] = 0;
3374
3375 pcpu_setup_first_chunk(ai, fc);
3376 pcpu_free_alloc_info(ai);
3377}
3378
3379#endif /* CONFIG_SMP */
3380
3381/*
3382 * pcpu_nr_pages - calculate total number of populated backing pages
3383 *
3384 * This reflects the number of pages populated to back chunks. Metadata is
3385 * excluded in the number exposed in meminfo as the number of backing pages
3386 * scales with the number of cpus and can quickly outweigh the memory used for
3387 * metadata. It also keeps this calculation nice and simple.
3388 *
3389 * RETURNS:
3390 * Total number of populated backing pages in use by the allocator.
3391 */
3392unsigned long pcpu_nr_pages(void)
3393{
3394 return pcpu_nr_populated * pcpu_nr_units;
3395}
3396
3397/*
3398 * Percpu allocator is initialized early during boot when neither slab or
3399 * workqueue is available. Plug async management until everything is up
3400 * and running.
3401 */
3402static int __init percpu_enable_async(void)
3403{
3404 pcpu_async_enabled = true;
3405 return 0;
3406}
3407subsys_initcall(percpu_enable_async);
1/*
2 * mm/percpu.c - percpu memory allocator
3 *
4 * Copyright (C) 2009 SUSE Linux Products GmbH
5 * Copyright (C) 2009 Tejun Heo <tj@kernel.org>
6 *
7 * This file is released under the GPLv2.
8 *
9 * This is percpu allocator which can handle both static and dynamic
10 * areas. Percpu areas are allocated in chunks. Each chunk is
11 * consisted of boot-time determined number of units and the first
12 * chunk is used for static percpu variables in the kernel image
13 * (special boot time alloc/init handling necessary as these areas
14 * need to be brought up before allocation services are running).
15 * Unit grows as necessary and all units grow or shrink in unison.
16 * When a chunk is filled up, another chunk is allocated.
17 *
18 * c0 c1 c2
19 * ------------------- ------------------- ------------
20 * | u0 | u1 | u2 | u3 | | u0 | u1 | u2 | u3 | | u0 | u1 | u
21 * ------------------- ...... ------------------- .... ------------
22 *
23 * Allocation is done in offset-size areas of single unit space. Ie,
24 * an area of 512 bytes at 6k in c1 occupies 512 bytes at 6k of c1:u0,
25 * c1:u1, c1:u2 and c1:u3. On UMA, units corresponds directly to
26 * cpus. On NUMA, the mapping can be non-linear and even sparse.
27 * Percpu access can be done by configuring percpu base registers
28 * according to cpu to unit mapping and pcpu_unit_size.
29 *
30 * There are usually many small percpu allocations many of them being
31 * as small as 4 bytes. The allocator organizes chunks into lists
32 * according to free size and tries to allocate from the fullest one.
33 * Each chunk keeps the maximum contiguous area size hint which is
34 * guaranteed to be equal to or larger than the maximum contiguous
35 * area in the chunk. This helps the allocator not to iterate the
36 * chunk maps unnecessarily.
37 *
38 * Allocation state in each chunk is kept using an array of integers
39 * on chunk->map. A positive value in the map represents a free
40 * region and negative allocated. Allocation inside a chunk is done
41 * by scanning this map sequentially and serving the first matching
42 * entry. This is mostly copied from the percpu_modalloc() allocator.
43 * Chunks can be determined from the address using the index field
44 * in the page struct. The index field contains a pointer to the chunk.
45 *
46 * To use this allocator, arch code should do the followings.
47 *
48 * - define __addr_to_pcpu_ptr() and __pcpu_ptr_to_addr() to translate
49 * regular address to percpu pointer and back if they need to be
50 * different from the default
51 *
52 * - use pcpu_setup_first_chunk() during percpu area initialization to
53 * setup the first chunk containing the kernel static percpu area
54 */
55
56#include <linux/bitmap.h>
57#include <linux/bootmem.h>
58#include <linux/err.h>
59#include <linux/list.h>
60#include <linux/log2.h>
61#include <linux/mm.h>
62#include <linux/module.h>
63#include <linux/mutex.h>
64#include <linux/percpu.h>
65#include <linux/pfn.h>
66#include <linux/slab.h>
67#include <linux/spinlock.h>
68#include <linux/vmalloc.h>
69#include <linux/workqueue.h>
70#include <linux/kmemleak.h>
71
72#include <asm/cacheflush.h>
73#include <asm/sections.h>
74#include <asm/tlbflush.h>
75#include <asm/io.h>
76
77#define PCPU_SLOT_BASE_SHIFT 5 /* 1-31 shares the same slot */
78#define PCPU_DFL_MAP_ALLOC 16 /* start a map with 16 ents */
79
80#ifdef CONFIG_SMP
81/* default addr <-> pcpu_ptr mapping, override in asm/percpu.h if necessary */
82#ifndef __addr_to_pcpu_ptr
83#define __addr_to_pcpu_ptr(addr) \
84 (void __percpu *)((unsigned long)(addr) - \
85 (unsigned long)pcpu_base_addr + \
86 (unsigned long)__per_cpu_start)
87#endif
88#ifndef __pcpu_ptr_to_addr
89#define __pcpu_ptr_to_addr(ptr) \
90 (void __force *)((unsigned long)(ptr) + \
91 (unsigned long)pcpu_base_addr - \
92 (unsigned long)__per_cpu_start)
93#endif
94#else /* CONFIG_SMP */
95/* on UP, it's always identity mapped */
96#define __addr_to_pcpu_ptr(addr) (void __percpu *)(addr)
97#define __pcpu_ptr_to_addr(ptr) (void __force *)(ptr)
98#endif /* CONFIG_SMP */
99
100struct pcpu_chunk {
101 struct list_head list; /* linked to pcpu_slot lists */
102 int free_size; /* free bytes in the chunk */
103 int contig_hint; /* max contiguous size hint */
104 void *base_addr; /* base address of this chunk */
105 int map_used; /* # of map entries used before the sentry */
106 int map_alloc; /* # of map entries allocated */
107 int *map; /* allocation map */
108 void *data; /* chunk data */
109 int first_free; /* no free below this */
110 bool immutable; /* no [de]population allowed */
111 unsigned long populated[]; /* populated bitmap */
112};
113
114static int pcpu_unit_pages __read_mostly;
115static int pcpu_unit_size __read_mostly;
116static int pcpu_nr_units __read_mostly;
117static int pcpu_atom_size __read_mostly;
118static int pcpu_nr_slots __read_mostly;
119static size_t pcpu_chunk_struct_size __read_mostly;
120
121/* cpus with the lowest and highest unit addresses */
122static unsigned int pcpu_low_unit_cpu __read_mostly;
123static unsigned int pcpu_high_unit_cpu __read_mostly;
124
125/* the address of the first chunk which starts with the kernel static area */
126void *pcpu_base_addr __read_mostly;
127EXPORT_SYMBOL_GPL(pcpu_base_addr);
128
129static const int *pcpu_unit_map __read_mostly; /* cpu -> unit */
130const unsigned long *pcpu_unit_offsets __read_mostly; /* cpu -> unit offset */
131
132/* group information, used for vm allocation */
133static int pcpu_nr_groups __read_mostly;
134static const unsigned long *pcpu_group_offsets __read_mostly;
135static const size_t *pcpu_group_sizes __read_mostly;
136
137/*
138 * The first chunk which always exists. Note that unlike other
139 * chunks, this one can be allocated and mapped in several different
140 * ways and thus often doesn't live in the vmalloc area.
141 */
142static struct pcpu_chunk *pcpu_first_chunk;
143
144/*
145 * Optional reserved chunk. This chunk reserves part of the first
146 * chunk and serves it for reserved allocations. The amount of
147 * reserved offset is in pcpu_reserved_chunk_limit. When reserved
148 * area doesn't exist, the following variables contain NULL and 0
149 * respectively.
150 */
151static struct pcpu_chunk *pcpu_reserved_chunk;
152static int pcpu_reserved_chunk_limit;
153
154/*
155 * Synchronization rules.
156 *
157 * There are two locks - pcpu_alloc_mutex and pcpu_lock. The former
158 * protects allocation/reclaim paths, chunks, populated bitmap and
159 * vmalloc mapping. The latter is a spinlock and protects the index
160 * data structures - chunk slots, chunks and area maps in chunks.
161 *
162 * During allocation, pcpu_alloc_mutex is kept locked all the time and
163 * pcpu_lock is grabbed and released as necessary. All actual memory
164 * allocations are done using GFP_KERNEL with pcpu_lock released. In
165 * general, percpu memory can't be allocated with irq off but
166 * irqsave/restore are still used in alloc path so that it can be used
167 * from early init path - sched_init() specifically.
168 *
169 * Free path accesses and alters only the index data structures, so it
170 * can be safely called from atomic context. When memory needs to be
171 * returned to the system, free path schedules reclaim_work which
172 * grabs both pcpu_alloc_mutex and pcpu_lock, unlinks chunks to be
173 * reclaimed, release both locks and frees the chunks. Note that it's
174 * necessary to grab both locks to remove a chunk from circulation as
175 * allocation path might be referencing the chunk with only
176 * pcpu_alloc_mutex locked.
177 */
178static DEFINE_MUTEX(pcpu_alloc_mutex); /* protects whole alloc and reclaim */
179static DEFINE_SPINLOCK(pcpu_lock); /* protects index data structures */
180
181static struct list_head *pcpu_slot __read_mostly; /* chunk list slots */
182
183/* reclaim work to release fully free chunks, scheduled from free path */
184static void pcpu_reclaim(struct work_struct *work);
185static DECLARE_WORK(pcpu_reclaim_work, pcpu_reclaim);
186
187static bool pcpu_addr_in_first_chunk(void *addr)
188{
189 void *first_start = pcpu_first_chunk->base_addr;
190
191 return addr >= first_start && addr < first_start + pcpu_unit_size;
192}
193
194static bool pcpu_addr_in_reserved_chunk(void *addr)
195{
196 void *first_start = pcpu_first_chunk->base_addr;
197
198 return addr >= first_start &&
199 addr < first_start + pcpu_reserved_chunk_limit;
200}
201
202static int __pcpu_size_to_slot(int size)
203{
204 int highbit = fls(size); /* size is in bytes */
205 return max(highbit - PCPU_SLOT_BASE_SHIFT + 2, 1);
206}
207
208static int pcpu_size_to_slot(int size)
209{
210 if (size == pcpu_unit_size)
211 return pcpu_nr_slots - 1;
212 return __pcpu_size_to_slot(size);
213}
214
215static int pcpu_chunk_slot(const struct pcpu_chunk *chunk)
216{
217 if (chunk->free_size < sizeof(int) || chunk->contig_hint < sizeof(int))
218 return 0;
219
220 return pcpu_size_to_slot(chunk->free_size);
221}
222
223/* set the pointer to a chunk in a page struct */
224static void pcpu_set_page_chunk(struct page *page, struct pcpu_chunk *pcpu)
225{
226 page->index = (unsigned long)pcpu;
227}
228
229/* obtain pointer to a chunk from a page struct */
230static struct pcpu_chunk *pcpu_get_page_chunk(struct page *page)
231{
232 return (struct pcpu_chunk *)page->index;
233}
234
235static int __maybe_unused pcpu_page_idx(unsigned int cpu, int page_idx)
236{
237 return pcpu_unit_map[cpu] * pcpu_unit_pages + page_idx;
238}
239
240static unsigned long pcpu_chunk_addr(struct pcpu_chunk *chunk,
241 unsigned int cpu, int page_idx)
242{
243 return (unsigned long)chunk->base_addr + pcpu_unit_offsets[cpu] +
244 (page_idx << PAGE_SHIFT);
245}
246
247static void __maybe_unused pcpu_next_unpop(struct pcpu_chunk *chunk,
248 int *rs, int *re, int end)
249{
250 *rs = find_next_zero_bit(chunk->populated, end, *rs);
251 *re = find_next_bit(chunk->populated, end, *rs + 1);
252}
253
254static void __maybe_unused pcpu_next_pop(struct pcpu_chunk *chunk,
255 int *rs, int *re, int end)
256{
257 *rs = find_next_bit(chunk->populated, end, *rs);
258 *re = find_next_zero_bit(chunk->populated, end, *rs + 1);
259}
260
261/*
262 * (Un)populated page region iterators. Iterate over (un)populated
263 * page regions between @start and @end in @chunk. @rs and @re should
264 * be integer variables and will be set to start and end page index of
265 * the current region.
266 */
267#define pcpu_for_each_unpop_region(chunk, rs, re, start, end) \
268 for ((rs) = (start), pcpu_next_unpop((chunk), &(rs), &(re), (end)); \
269 (rs) < (re); \
270 (rs) = (re) + 1, pcpu_next_unpop((chunk), &(rs), &(re), (end)))
271
272#define pcpu_for_each_pop_region(chunk, rs, re, start, end) \
273 for ((rs) = (start), pcpu_next_pop((chunk), &(rs), &(re), (end)); \
274 (rs) < (re); \
275 (rs) = (re) + 1, pcpu_next_pop((chunk), &(rs), &(re), (end)))
276
277/**
278 * pcpu_mem_zalloc - allocate memory
279 * @size: bytes to allocate
280 *
281 * Allocate @size bytes. If @size is smaller than PAGE_SIZE,
282 * kzalloc() is used; otherwise, vzalloc() is used. The returned
283 * memory is always zeroed.
284 *
285 * CONTEXT:
286 * Does GFP_KERNEL allocation.
287 *
288 * RETURNS:
289 * Pointer to the allocated area on success, NULL on failure.
290 */
291static void *pcpu_mem_zalloc(size_t size)
292{
293 if (WARN_ON_ONCE(!slab_is_available()))
294 return NULL;
295
296 if (size <= PAGE_SIZE)
297 return kzalloc(size, GFP_KERNEL);
298 else
299 return vzalloc(size);
300}
301
302/**
303 * pcpu_mem_free - free memory
304 * @ptr: memory to free
305 * @size: size of the area
306 *
307 * Free @ptr. @ptr should have been allocated using pcpu_mem_zalloc().
308 */
309static void pcpu_mem_free(void *ptr, size_t size)
310{
311 if (size <= PAGE_SIZE)
312 kfree(ptr);
313 else
314 vfree(ptr);
315}
316
317/**
318 * pcpu_chunk_relocate - put chunk in the appropriate chunk slot
319 * @chunk: chunk of interest
320 * @oslot: the previous slot it was on
321 *
322 * This function is called after an allocation or free changed @chunk.
323 * New slot according to the changed state is determined and @chunk is
324 * moved to the slot. Note that the reserved chunk is never put on
325 * chunk slots.
326 *
327 * CONTEXT:
328 * pcpu_lock.
329 */
330static void pcpu_chunk_relocate(struct pcpu_chunk *chunk, int oslot)
331{
332 int nslot = pcpu_chunk_slot(chunk);
333
334 if (chunk != pcpu_reserved_chunk && oslot != nslot) {
335 if (oslot < nslot)
336 list_move(&chunk->list, &pcpu_slot[nslot]);
337 else
338 list_move_tail(&chunk->list, &pcpu_slot[nslot]);
339 }
340}
341
342/**
343 * pcpu_need_to_extend - determine whether chunk area map needs to be extended
344 * @chunk: chunk of interest
345 *
346 * Determine whether area map of @chunk needs to be extended to
347 * accommodate a new allocation.
348 *
349 * CONTEXT:
350 * pcpu_lock.
351 *
352 * RETURNS:
353 * New target map allocation length if extension is necessary, 0
354 * otherwise.
355 */
356static int pcpu_need_to_extend(struct pcpu_chunk *chunk)
357{
358 int new_alloc;
359
360 if (chunk->map_alloc >= chunk->map_used + 3)
361 return 0;
362
363 new_alloc = PCPU_DFL_MAP_ALLOC;
364 while (new_alloc < chunk->map_used + 3)
365 new_alloc *= 2;
366
367 return new_alloc;
368}
369
370/**
371 * pcpu_extend_area_map - extend area map of a chunk
372 * @chunk: chunk of interest
373 * @new_alloc: new target allocation length of the area map
374 *
375 * Extend area map of @chunk to have @new_alloc entries.
376 *
377 * CONTEXT:
378 * Does GFP_KERNEL allocation. Grabs and releases pcpu_lock.
379 *
380 * RETURNS:
381 * 0 on success, -errno on failure.
382 */
383static int pcpu_extend_area_map(struct pcpu_chunk *chunk, int new_alloc)
384{
385 int *old = NULL, *new = NULL;
386 size_t old_size = 0, new_size = new_alloc * sizeof(new[0]);
387 unsigned long flags;
388
389 new = pcpu_mem_zalloc(new_size);
390 if (!new)
391 return -ENOMEM;
392
393 /* acquire pcpu_lock and switch to new area map */
394 spin_lock_irqsave(&pcpu_lock, flags);
395
396 if (new_alloc <= chunk->map_alloc)
397 goto out_unlock;
398
399 old_size = chunk->map_alloc * sizeof(chunk->map[0]);
400 old = chunk->map;
401
402 memcpy(new, old, old_size);
403
404 chunk->map_alloc = new_alloc;
405 chunk->map = new;
406 new = NULL;
407
408out_unlock:
409 spin_unlock_irqrestore(&pcpu_lock, flags);
410
411 /*
412 * pcpu_mem_free() might end up calling vfree() which uses
413 * IRQ-unsafe lock and thus can't be called under pcpu_lock.
414 */
415 pcpu_mem_free(old, old_size);
416 pcpu_mem_free(new, new_size);
417
418 return 0;
419}
420
421/**
422 * pcpu_alloc_area - allocate area from a pcpu_chunk
423 * @chunk: chunk of interest
424 * @size: wanted size in bytes
425 * @align: wanted align
426 *
427 * Try to allocate @size bytes area aligned at @align from @chunk.
428 * Note that this function only allocates the offset. It doesn't
429 * populate or map the area.
430 *
431 * @chunk->map must have at least two free slots.
432 *
433 * CONTEXT:
434 * pcpu_lock.
435 *
436 * RETURNS:
437 * Allocated offset in @chunk on success, -1 if no matching area is
438 * found.
439 */
440static int pcpu_alloc_area(struct pcpu_chunk *chunk, int size, int align)
441{
442 int oslot = pcpu_chunk_slot(chunk);
443 int max_contig = 0;
444 int i, off;
445 bool seen_free = false;
446 int *p;
447
448 for (i = chunk->first_free, p = chunk->map + i; i < chunk->map_used; i++, p++) {
449 int head, tail;
450 int this_size;
451
452 off = *p;
453 if (off & 1)
454 continue;
455
456 /* extra for alignment requirement */
457 head = ALIGN(off, align) - off;
458
459 this_size = (p[1] & ~1) - off;
460 if (this_size < head + size) {
461 if (!seen_free) {
462 chunk->first_free = i;
463 seen_free = true;
464 }
465 max_contig = max(this_size, max_contig);
466 continue;
467 }
468
469 /*
470 * If head is small or the previous block is free,
471 * merge'em. Note that 'small' is defined as smaller
472 * than sizeof(int), which is very small but isn't too
473 * uncommon for percpu allocations.
474 */
475 if (head && (head < sizeof(int) || !(p[-1] & 1))) {
476 *p = off += head;
477 if (p[-1] & 1)
478 chunk->free_size -= head;
479 else
480 max_contig = max(*p - p[-1], max_contig);
481 this_size -= head;
482 head = 0;
483 }
484
485 /* if tail is small, just keep it around */
486 tail = this_size - head - size;
487 if (tail < sizeof(int)) {
488 tail = 0;
489 size = this_size - head;
490 }
491
492 /* split if warranted */
493 if (head || tail) {
494 int nr_extra = !!head + !!tail;
495
496 /* insert new subblocks */
497 memmove(p + nr_extra + 1, p + 1,
498 sizeof(chunk->map[0]) * (chunk->map_used - i));
499 chunk->map_used += nr_extra;
500
501 if (head) {
502 if (!seen_free) {
503 chunk->first_free = i;
504 seen_free = true;
505 }
506 *++p = off += head;
507 ++i;
508 max_contig = max(head, max_contig);
509 }
510 if (tail) {
511 p[1] = off + size;
512 max_contig = max(tail, max_contig);
513 }
514 }
515
516 if (!seen_free)
517 chunk->first_free = i + 1;
518
519 /* update hint and mark allocated */
520 if (i + 1 == chunk->map_used)
521 chunk->contig_hint = max_contig; /* fully scanned */
522 else
523 chunk->contig_hint = max(chunk->contig_hint,
524 max_contig);
525
526 chunk->free_size -= size;
527 *p |= 1;
528
529 pcpu_chunk_relocate(chunk, oslot);
530 return off;
531 }
532
533 chunk->contig_hint = max_contig; /* fully scanned */
534 pcpu_chunk_relocate(chunk, oslot);
535
536 /* tell the upper layer that this chunk has no matching area */
537 return -1;
538}
539
540/**
541 * pcpu_free_area - free area to a pcpu_chunk
542 * @chunk: chunk of interest
543 * @freeme: offset of area to free
544 *
545 * Free area starting from @freeme to @chunk. Note that this function
546 * only modifies the allocation map. It doesn't depopulate or unmap
547 * the area.
548 *
549 * CONTEXT:
550 * pcpu_lock.
551 */
552static void pcpu_free_area(struct pcpu_chunk *chunk, int freeme)
553{
554 int oslot = pcpu_chunk_slot(chunk);
555 int off = 0;
556 unsigned i, j;
557 int to_free = 0;
558 int *p;
559
560 freeme |= 1; /* we are searching for <given offset, in use> pair */
561
562 i = 0;
563 j = chunk->map_used;
564 while (i != j) {
565 unsigned k = (i + j) / 2;
566 off = chunk->map[k];
567 if (off < freeme)
568 i = k + 1;
569 else if (off > freeme)
570 j = k;
571 else
572 i = j = k;
573 }
574 BUG_ON(off != freeme);
575
576 if (i < chunk->first_free)
577 chunk->first_free = i;
578
579 p = chunk->map + i;
580 *p = off &= ~1;
581 chunk->free_size += (p[1] & ~1) - off;
582
583 /* merge with next? */
584 if (!(p[1] & 1))
585 to_free++;
586 /* merge with previous? */
587 if (i > 0 && !(p[-1] & 1)) {
588 to_free++;
589 i--;
590 p--;
591 }
592 if (to_free) {
593 chunk->map_used -= to_free;
594 memmove(p + 1, p + 1 + to_free,
595 (chunk->map_used - i) * sizeof(chunk->map[0]));
596 }
597
598 chunk->contig_hint = max(chunk->map[i + 1] - chunk->map[i] - 1, chunk->contig_hint);
599 pcpu_chunk_relocate(chunk, oslot);
600}
601
602static struct pcpu_chunk *pcpu_alloc_chunk(void)
603{
604 struct pcpu_chunk *chunk;
605
606 chunk = pcpu_mem_zalloc(pcpu_chunk_struct_size);
607 if (!chunk)
608 return NULL;
609
610 chunk->map = pcpu_mem_zalloc(PCPU_DFL_MAP_ALLOC *
611 sizeof(chunk->map[0]));
612 if (!chunk->map) {
613 pcpu_mem_free(chunk, pcpu_chunk_struct_size);
614 return NULL;
615 }
616
617 chunk->map_alloc = PCPU_DFL_MAP_ALLOC;
618 chunk->map[0] = 0;
619 chunk->map[1] = pcpu_unit_size | 1;
620 chunk->map_used = 1;
621
622 INIT_LIST_HEAD(&chunk->list);
623 chunk->free_size = pcpu_unit_size;
624 chunk->contig_hint = pcpu_unit_size;
625
626 return chunk;
627}
628
629static void pcpu_free_chunk(struct pcpu_chunk *chunk)
630{
631 if (!chunk)
632 return;
633 pcpu_mem_free(chunk->map, chunk->map_alloc * sizeof(chunk->map[0]));
634 pcpu_mem_free(chunk, pcpu_chunk_struct_size);
635}
636
637/*
638 * Chunk management implementation.
639 *
640 * To allow different implementations, chunk alloc/free and
641 * [de]population are implemented in a separate file which is pulled
642 * into this file and compiled together. The following functions
643 * should be implemented.
644 *
645 * pcpu_populate_chunk - populate the specified range of a chunk
646 * pcpu_depopulate_chunk - depopulate the specified range of a chunk
647 * pcpu_create_chunk - create a new chunk
648 * pcpu_destroy_chunk - destroy a chunk, always preceded by full depop
649 * pcpu_addr_to_page - translate address to physical address
650 * pcpu_verify_alloc_info - check alloc_info is acceptable during init
651 */
652static int pcpu_populate_chunk(struct pcpu_chunk *chunk, int off, int size);
653static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk, int off, int size);
654static struct pcpu_chunk *pcpu_create_chunk(void);
655static void pcpu_destroy_chunk(struct pcpu_chunk *chunk);
656static struct page *pcpu_addr_to_page(void *addr);
657static int __init pcpu_verify_alloc_info(const struct pcpu_alloc_info *ai);
658
659#ifdef CONFIG_NEED_PER_CPU_KM
660#include "percpu-km.c"
661#else
662#include "percpu-vm.c"
663#endif
664
665/**
666 * pcpu_chunk_addr_search - determine chunk containing specified address
667 * @addr: address for which the chunk needs to be determined.
668 *
669 * RETURNS:
670 * The address of the found chunk.
671 */
672static struct pcpu_chunk *pcpu_chunk_addr_search(void *addr)
673{
674 /* is it in the first chunk? */
675 if (pcpu_addr_in_first_chunk(addr)) {
676 /* is it in the reserved area? */
677 if (pcpu_addr_in_reserved_chunk(addr))
678 return pcpu_reserved_chunk;
679 return pcpu_first_chunk;
680 }
681
682 /*
683 * The address is relative to unit0 which might be unused and
684 * thus unmapped. Offset the address to the unit space of the
685 * current processor before looking it up in the vmalloc
686 * space. Note that any possible cpu id can be used here, so
687 * there's no need to worry about preemption or cpu hotplug.
688 */
689 addr += pcpu_unit_offsets[raw_smp_processor_id()];
690 return pcpu_get_page_chunk(pcpu_addr_to_page(addr));
691}
692
693/**
694 * pcpu_alloc - the percpu allocator
695 * @size: size of area to allocate in bytes
696 * @align: alignment of area (max PAGE_SIZE)
697 * @reserved: allocate from the reserved chunk if available
698 *
699 * Allocate percpu area of @size bytes aligned at @align.
700 *
701 * CONTEXT:
702 * Does GFP_KERNEL allocation.
703 *
704 * RETURNS:
705 * Percpu pointer to the allocated area on success, NULL on failure.
706 */
707static void __percpu *pcpu_alloc(size_t size, size_t align, bool reserved)
708{
709 static int warn_limit = 10;
710 struct pcpu_chunk *chunk;
711 const char *err;
712 int slot, off, new_alloc;
713 unsigned long flags;
714 void __percpu *ptr;
715
716 /*
717 * We want the lowest bit of offset available for in-use/free
718 * indicator, so force >= 16bit alignment and make size even.
719 */
720 if (unlikely(align < 2))
721 align = 2;
722
723 if (unlikely(size & 1))
724 size++;
725
726 if (unlikely(!size || size > PCPU_MIN_UNIT_SIZE || align > PAGE_SIZE)) {
727 WARN(true, "illegal size (%zu) or align (%zu) for "
728 "percpu allocation\n", size, align);
729 return NULL;
730 }
731
732 mutex_lock(&pcpu_alloc_mutex);
733 spin_lock_irqsave(&pcpu_lock, flags);
734
735 /* serve reserved allocations from the reserved chunk if available */
736 if (reserved && pcpu_reserved_chunk) {
737 chunk = pcpu_reserved_chunk;
738
739 if (size > chunk->contig_hint) {
740 err = "alloc from reserved chunk failed";
741 goto fail_unlock;
742 }
743
744 while ((new_alloc = pcpu_need_to_extend(chunk))) {
745 spin_unlock_irqrestore(&pcpu_lock, flags);
746 if (pcpu_extend_area_map(chunk, new_alloc) < 0) {
747 err = "failed to extend area map of reserved chunk";
748 goto fail_unlock_mutex;
749 }
750 spin_lock_irqsave(&pcpu_lock, flags);
751 }
752
753 off = pcpu_alloc_area(chunk, size, align);
754 if (off >= 0)
755 goto area_found;
756
757 err = "alloc from reserved chunk failed";
758 goto fail_unlock;
759 }
760
761restart:
762 /* search through normal chunks */
763 for (slot = pcpu_size_to_slot(size); slot < pcpu_nr_slots; slot++) {
764 list_for_each_entry(chunk, &pcpu_slot[slot], list) {
765 if (size > chunk->contig_hint)
766 continue;
767
768 new_alloc = pcpu_need_to_extend(chunk);
769 if (new_alloc) {
770 spin_unlock_irqrestore(&pcpu_lock, flags);
771 if (pcpu_extend_area_map(chunk,
772 new_alloc) < 0) {
773 err = "failed to extend area map";
774 goto fail_unlock_mutex;
775 }
776 spin_lock_irqsave(&pcpu_lock, flags);
777 /*
778 * pcpu_lock has been dropped, need to
779 * restart cpu_slot list walking.
780 */
781 goto restart;
782 }
783
784 off = pcpu_alloc_area(chunk, size, align);
785 if (off >= 0)
786 goto area_found;
787 }
788 }
789
790 /* hmmm... no space left, create a new chunk */
791 spin_unlock_irqrestore(&pcpu_lock, flags);
792
793 chunk = pcpu_create_chunk();
794 if (!chunk) {
795 err = "failed to allocate new chunk";
796 goto fail_unlock_mutex;
797 }
798
799 spin_lock_irqsave(&pcpu_lock, flags);
800 pcpu_chunk_relocate(chunk, -1);
801 goto restart;
802
803area_found:
804 spin_unlock_irqrestore(&pcpu_lock, flags);
805
806 /* populate, map and clear the area */
807 if (pcpu_populate_chunk(chunk, off, size)) {
808 spin_lock_irqsave(&pcpu_lock, flags);
809 pcpu_free_area(chunk, off);
810 err = "failed to populate";
811 goto fail_unlock;
812 }
813
814 mutex_unlock(&pcpu_alloc_mutex);
815
816 /* return address relative to base address */
817 ptr = __addr_to_pcpu_ptr(chunk->base_addr + off);
818 kmemleak_alloc_percpu(ptr, size);
819 return ptr;
820
821fail_unlock:
822 spin_unlock_irqrestore(&pcpu_lock, flags);
823fail_unlock_mutex:
824 mutex_unlock(&pcpu_alloc_mutex);
825 if (warn_limit) {
826 pr_warning("PERCPU: allocation failed, size=%zu align=%zu, "
827 "%s\n", size, align, err);
828 dump_stack();
829 if (!--warn_limit)
830 pr_info("PERCPU: limit reached, disable warning\n");
831 }
832 return NULL;
833}
834
835/**
836 * __alloc_percpu - allocate dynamic percpu area
837 * @size: size of area to allocate in bytes
838 * @align: alignment of area (max PAGE_SIZE)
839 *
840 * Allocate zero-filled percpu area of @size bytes aligned at @align.
841 * Might sleep. Might trigger writeouts.
842 *
843 * CONTEXT:
844 * Does GFP_KERNEL allocation.
845 *
846 * RETURNS:
847 * Percpu pointer to the allocated area on success, NULL on failure.
848 */
849void __percpu *__alloc_percpu(size_t size, size_t align)
850{
851 return pcpu_alloc(size, align, false);
852}
853EXPORT_SYMBOL_GPL(__alloc_percpu);
854
855/**
856 * __alloc_reserved_percpu - allocate reserved percpu area
857 * @size: size of area to allocate in bytes
858 * @align: alignment of area (max PAGE_SIZE)
859 *
860 * Allocate zero-filled percpu area of @size bytes aligned at @align
861 * from reserved percpu area if arch has set it up; otherwise,
862 * allocation is served from the same dynamic area. Might sleep.
863 * Might trigger writeouts.
864 *
865 * CONTEXT:
866 * Does GFP_KERNEL allocation.
867 *
868 * RETURNS:
869 * Percpu pointer to the allocated area on success, NULL on failure.
870 */
871void __percpu *__alloc_reserved_percpu(size_t size, size_t align)
872{
873 return pcpu_alloc(size, align, true);
874}
875
876/**
877 * pcpu_reclaim - reclaim fully free chunks, workqueue function
878 * @work: unused
879 *
880 * Reclaim all fully free chunks except for the first one.
881 *
882 * CONTEXT:
883 * workqueue context.
884 */
885static void pcpu_reclaim(struct work_struct *work)
886{
887 LIST_HEAD(todo);
888 struct list_head *head = &pcpu_slot[pcpu_nr_slots - 1];
889 struct pcpu_chunk *chunk, *next;
890
891 mutex_lock(&pcpu_alloc_mutex);
892 spin_lock_irq(&pcpu_lock);
893
894 list_for_each_entry_safe(chunk, next, head, list) {
895 WARN_ON(chunk->immutable);
896
897 /* spare the first one */
898 if (chunk == list_first_entry(head, struct pcpu_chunk, list))
899 continue;
900
901 list_move(&chunk->list, &todo);
902 }
903
904 spin_unlock_irq(&pcpu_lock);
905
906 list_for_each_entry_safe(chunk, next, &todo, list) {
907 pcpu_depopulate_chunk(chunk, 0, pcpu_unit_size);
908 pcpu_destroy_chunk(chunk);
909 }
910
911 mutex_unlock(&pcpu_alloc_mutex);
912}
913
914/**
915 * free_percpu - free percpu area
916 * @ptr: pointer to area to free
917 *
918 * Free percpu area @ptr.
919 *
920 * CONTEXT:
921 * Can be called from atomic context.
922 */
923void free_percpu(void __percpu *ptr)
924{
925 void *addr;
926 struct pcpu_chunk *chunk;
927 unsigned long flags;
928 int off;
929
930 if (!ptr)
931 return;
932
933 kmemleak_free_percpu(ptr);
934
935 addr = __pcpu_ptr_to_addr(ptr);
936
937 spin_lock_irqsave(&pcpu_lock, flags);
938
939 chunk = pcpu_chunk_addr_search(addr);
940 off = addr - chunk->base_addr;
941
942 pcpu_free_area(chunk, off);
943
944 /* if there are more than one fully free chunks, wake up grim reaper */
945 if (chunk->free_size == pcpu_unit_size) {
946 struct pcpu_chunk *pos;
947
948 list_for_each_entry(pos, &pcpu_slot[pcpu_nr_slots - 1], list)
949 if (pos != chunk) {
950 schedule_work(&pcpu_reclaim_work);
951 break;
952 }
953 }
954
955 spin_unlock_irqrestore(&pcpu_lock, flags);
956}
957EXPORT_SYMBOL_GPL(free_percpu);
958
959/**
960 * is_kernel_percpu_address - test whether address is from static percpu area
961 * @addr: address to test
962 *
963 * Test whether @addr belongs to in-kernel static percpu area. Module
964 * static percpu areas are not considered. For those, use
965 * is_module_percpu_address().
966 *
967 * RETURNS:
968 * %true if @addr is from in-kernel static percpu area, %false otherwise.
969 */
970bool is_kernel_percpu_address(unsigned long addr)
971{
972#ifdef CONFIG_SMP
973 const size_t static_size = __per_cpu_end - __per_cpu_start;
974 void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr);
975 unsigned int cpu;
976
977 for_each_possible_cpu(cpu) {
978 void *start = per_cpu_ptr(base, cpu);
979
980 if ((void *)addr >= start && (void *)addr < start + static_size)
981 return true;
982 }
983#endif
984 /* on UP, can't distinguish from other static vars, always false */
985 return false;
986}
987
988/**
989 * per_cpu_ptr_to_phys - convert translated percpu address to physical address
990 * @addr: the address to be converted to physical address
991 *
992 * Given @addr which is dereferenceable address obtained via one of
993 * percpu access macros, this function translates it into its physical
994 * address. The caller is responsible for ensuring @addr stays valid
995 * until this function finishes.
996 *
997 * percpu allocator has special setup for the first chunk, which currently
998 * supports either embedding in linear address space or vmalloc mapping,
999 * and, from the second one, the backing allocator (currently either vm or
1000 * km) provides translation.
1001 *
1002 * The addr can be tranlated simply without checking if it falls into the
1003 * first chunk. But the current code reflects better how percpu allocator
1004 * actually works, and the verification can discover both bugs in percpu
1005 * allocator itself and per_cpu_ptr_to_phys() callers. So we keep current
1006 * code.
1007 *
1008 * RETURNS:
1009 * The physical address for @addr.
1010 */
1011phys_addr_t per_cpu_ptr_to_phys(void *addr)
1012{
1013 void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr);
1014 bool in_first_chunk = false;
1015 unsigned long first_low, first_high;
1016 unsigned int cpu;
1017
1018 /*
1019 * The following test on unit_low/high isn't strictly
1020 * necessary but will speed up lookups of addresses which
1021 * aren't in the first chunk.
1022 */
1023 first_low = pcpu_chunk_addr(pcpu_first_chunk, pcpu_low_unit_cpu, 0);
1024 first_high = pcpu_chunk_addr(pcpu_first_chunk, pcpu_high_unit_cpu,
1025 pcpu_unit_pages);
1026 if ((unsigned long)addr >= first_low &&
1027 (unsigned long)addr < first_high) {
1028 for_each_possible_cpu(cpu) {
1029 void *start = per_cpu_ptr(base, cpu);
1030
1031 if (addr >= start && addr < start + pcpu_unit_size) {
1032 in_first_chunk = true;
1033 break;
1034 }
1035 }
1036 }
1037
1038 if (in_first_chunk) {
1039 if (!is_vmalloc_addr(addr))
1040 return __pa(addr);
1041 else
1042 return page_to_phys(vmalloc_to_page(addr)) +
1043 offset_in_page(addr);
1044 } else
1045 return page_to_phys(pcpu_addr_to_page(addr)) +
1046 offset_in_page(addr);
1047}
1048
1049/**
1050 * pcpu_alloc_alloc_info - allocate percpu allocation info
1051 * @nr_groups: the number of groups
1052 * @nr_units: the number of units
1053 *
1054 * Allocate ai which is large enough for @nr_groups groups containing
1055 * @nr_units units. The returned ai's groups[0].cpu_map points to the
1056 * cpu_map array which is long enough for @nr_units and filled with
1057 * NR_CPUS. It's the caller's responsibility to initialize cpu_map
1058 * pointer of other groups.
1059 *
1060 * RETURNS:
1061 * Pointer to the allocated pcpu_alloc_info on success, NULL on
1062 * failure.
1063 */
1064struct pcpu_alloc_info * __init pcpu_alloc_alloc_info(int nr_groups,
1065 int nr_units)
1066{
1067 struct pcpu_alloc_info *ai;
1068 size_t base_size, ai_size;
1069 void *ptr;
1070 int unit;
1071
1072 base_size = ALIGN(sizeof(*ai) + nr_groups * sizeof(ai->groups[0]),
1073 __alignof__(ai->groups[0].cpu_map[0]));
1074 ai_size = base_size + nr_units * sizeof(ai->groups[0].cpu_map[0]);
1075
1076 ptr = memblock_virt_alloc_nopanic(PFN_ALIGN(ai_size), 0);
1077 if (!ptr)
1078 return NULL;
1079 ai = ptr;
1080 ptr += base_size;
1081
1082 ai->groups[0].cpu_map = ptr;
1083
1084 for (unit = 0; unit < nr_units; unit++)
1085 ai->groups[0].cpu_map[unit] = NR_CPUS;
1086
1087 ai->nr_groups = nr_groups;
1088 ai->__ai_size = PFN_ALIGN(ai_size);
1089
1090 return ai;
1091}
1092
1093/**
1094 * pcpu_free_alloc_info - free percpu allocation info
1095 * @ai: pcpu_alloc_info to free
1096 *
1097 * Free @ai which was allocated by pcpu_alloc_alloc_info().
1098 */
1099void __init pcpu_free_alloc_info(struct pcpu_alloc_info *ai)
1100{
1101 memblock_free_early(__pa(ai), ai->__ai_size);
1102}
1103
1104/**
1105 * pcpu_dump_alloc_info - print out information about pcpu_alloc_info
1106 * @lvl: loglevel
1107 * @ai: allocation info to dump
1108 *
1109 * Print out information about @ai using loglevel @lvl.
1110 */
1111static void pcpu_dump_alloc_info(const char *lvl,
1112 const struct pcpu_alloc_info *ai)
1113{
1114 int group_width = 1, cpu_width = 1, width;
1115 char empty_str[] = "--------";
1116 int alloc = 0, alloc_end = 0;
1117 int group, v;
1118 int upa, apl; /* units per alloc, allocs per line */
1119
1120 v = ai->nr_groups;
1121 while (v /= 10)
1122 group_width++;
1123
1124 v = num_possible_cpus();
1125 while (v /= 10)
1126 cpu_width++;
1127 empty_str[min_t(int, cpu_width, sizeof(empty_str) - 1)] = '\0';
1128
1129 upa = ai->alloc_size / ai->unit_size;
1130 width = upa * (cpu_width + 1) + group_width + 3;
1131 apl = rounddown_pow_of_two(max(60 / width, 1));
1132
1133 printk("%spcpu-alloc: s%zu r%zu d%zu u%zu alloc=%zu*%zu",
1134 lvl, ai->static_size, ai->reserved_size, ai->dyn_size,
1135 ai->unit_size, ai->alloc_size / ai->atom_size, ai->atom_size);
1136
1137 for (group = 0; group < ai->nr_groups; group++) {
1138 const struct pcpu_group_info *gi = &ai->groups[group];
1139 int unit = 0, unit_end = 0;
1140
1141 BUG_ON(gi->nr_units % upa);
1142 for (alloc_end += gi->nr_units / upa;
1143 alloc < alloc_end; alloc++) {
1144 if (!(alloc % apl)) {
1145 printk(KERN_CONT "\n");
1146 printk("%spcpu-alloc: ", lvl);
1147 }
1148 printk(KERN_CONT "[%0*d] ", group_width, group);
1149
1150 for (unit_end += upa; unit < unit_end; unit++)
1151 if (gi->cpu_map[unit] != NR_CPUS)
1152 printk(KERN_CONT "%0*d ", cpu_width,
1153 gi->cpu_map[unit]);
1154 else
1155 printk(KERN_CONT "%s ", empty_str);
1156 }
1157 }
1158 printk(KERN_CONT "\n");
1159}
1160
1161/**
1162 * pcpu_setup_first_chunk - initialize the first percpu chunk
1163 * @ai: pcpu_alloc_info describing how to percpu area is shaped
1164 * @base_addr: mapped address
1165 *
1166 * Initialize the first percpu chunk which contains the kernel static
1167 * perpcu area. This function is to be called from arch percpu area
1168 * setup path.
1169 *
1170 * @ai contains all information necessary to initialize the first
1171 * chunk and prime the dynamic percpu allocator.
1172 *
1173 * @ai->static_size is the size of static percpu area.
1174 *
1175 * @ai->reserved_size, if non-zero, specifies the amount of bytes to
1176 * reserve after the static area in the first chunk. This reserves
1177 * the first chunk such that it's available only through reserved
1178 * percpu allocation. This is primarily used to serve module percpu
1179 * static areas on architectures where the addressing model has
1180 * limited offset range for symbol relocations to guarantee module
1181 * percpu symbols fall inside the relocatable range.
1182 *
1183 * @ai->dyn_size determines the number of bytes available for dynamic
1184 * allocation in the first chunk. The area between @ai->static_size +
1185 * @ai->reserved_size + @ai->dyn_size and @ai->unit_size is unused.
1186 *
1187 * @ai->unit_size specifies unit size and must be aligned to PAGE_SIZE
1188 * and equal to or larger than @ai->static_size + @ai->reserved_size +
1189 * @ai->dyn_size.
1190 *
1191 * @ai->atom_size is the allocation atom size and used as alignment
1192 * for vm areas.
1193 *
1194 * @ai->alloc_size is the allocation size and always multiple of
1195 * @ai->atom_size. This is larger than @ai->atom_size if
1196 * @ai->unit_size is larger than @ai->atom_size.
1197 *
1198 * @ai->nr_groups and @ai->groups describe virtual memory layout of
1199 * percpu areas. Units which should be colocated are put into the
1200 * same group. Dynamic VM areas will be allocated according to these
1201 * groupings. If @ai->nr_groups is zero, a single group containing
1202 * all units is assumed.
1203 *
1204 * The caller should have mapped the first chunk at @base_addr and
1205 * copied static data to each unit.
1206 *
1207 * If the first chunk ends up with both reserved and dynamic areas, it
1208 * is served by two chunks - one to serve the core static and reserved
1209 * areas and the other for the dynamic area. They share the same vm
1210 * and page map but uses different area allocation map to stay away
1211 * from each other. The latter chunk is circulated in the chunk slots
1212 * and available for dynamic allocation like any other chunks.
1213 *
1214 * RETURNS:
1215 * 0 on success, -errno on failure.
1216 */
1217int __init pcpu_setup_first_chunk(const struct pcpu_alloc_info *ai,
1218 void *base_addr)
1219{
1220 static char cpus_buf[4096] __initdata;
1221 static int smap[PERCPU_DYNAMIC_EARLY_SLOTS] __initdata;
1222 static int dmap[PERCPU_DYNAMIC_EARLY_SLOTS] __initdata;
1223 size_t dyn_size = ai->dyn_size;
1224 size_t size_sum = ai->static_size + ai->reserved_size + dyn_size;
1225 struct pcpu_chunk *schunk, *dchunk = NULL;
1226 unsigned long *group_offsets;
1227 size_t *group_sizes;
1228 unsigned long *unit_off;
1229 unsigned int cpu;
1230 int *unit_map;
1231 int group, unit, i;
1232
1233 cpumask_scnprintf(cpus_buf, sizeof(cpus_buf), cpu_possible_mask);
1234
1235#define PCPU_SETUP_BUG_ON(cond) do { \
1236 if (unlikely(cond)) { \
1237 pr_emerg("PERCPU: failed to initialize, %s", #cond); \
1238 pr_emerg("PERCPU: cpu_possible_mask=%s\n", cpus_buf); \
1239 pcpu_dump_alloc_info(KERN_EMERG, ai); \
1240 BUG(); \
1241 } \
1242} while (0)
1243
1244 /* sanity checks */
1245 PCPU_SETUP_BUG_ON(ai->nr_groups <= 0);
1246#ifdef CONFIG_SMP
1247 PCPU_SETUP_BUG_ON(!ai->static_size);
1248 PCPU_SETUP_BUG_ON((unsigned long)__per_cpu_start & ~PAGE_MASK);
1249#endif
1250 PCPU_SETUP_BUG_ON(!base_addr);
1251 PCPU_SETUP_BUG_ON((unsigned long)base_addr & ~PAGE_MASK);
1252 PCPU_SETUP_BUG_ON(ai->unit_size < size_sum);
1253 PCPU_SETUP_BUG_ON(ai->unit_size & ~PAGE_MASK);
1254 PCPU_SETUP_BUG_ON(ai->unit_size < PCPU_MIN_UNIT_SIZE);
1255 PCPU_SETUP_BUG_ON(ai->dyn_size < PERCPU_DYNAMIC_EARLY_SIZE);
1256 PCPU_SETUP_BUG_ON(pcpu_verify_alloc_info(ai) < 0);
1257
1258 /* process group information and build config tables accordingly */
1259 group_offsets = memblock_virt_alloc(ai->nr_groups *
1260 sizeof(group_offsets[0]), 0);
1261 group_sizes = memblock_virt_alloc(ai->nr_groups *
1262 sizeof(group_sizes[0]), 0);
1263 unit_map = memblock_virt_alloc(nr_cpu_ids * sizeof(unit_map[0]), 0);
1264 unit_off = memblock_virt_alloc(nr_cpu_ids * sizeof(unit_off[0]), 0);
1265
1266 for (cpu = 0; cpu < nr_cpu_ids; cpu++)
1267 unit_map[cpu] = UINT_MAX;
1268
1269 pcpu_low_unit_cpu = NR_CPUS;
1270 pcpu_high_unit_cpu = NR_CPUS;
1271
1272 for (group = 0, unit = 0; group < ai->nr_groups; group++, unit += i) {
1273 const struct pcpu_group_info *gi = &ai->groups[group];
1274
1275 group_offsets[group] = gi->base_offset;
1276 group_sizes[group] = gi->nr_units * ai->unit_size;
1277
1278 for (i = 0; i < gi->nr_units; i++) {
1279 cpu = gi->cpu_map[i];
1280 if (cpu == NR_CPUS)
1281 continue;
1282
1283 PCPU_SETUP_BUG_ON(cpu > nr_cpu_ids);
1284 PCPU_SETUP_BUG_ON(!cpu_possible(cpu));
1285 PCPU_SETUP_BUG_ON(unit_map[cpu] != UINT_MAX);
1286
1287 unit_map[cpu] = unit + i;
1288 unit_off[cpu] = gi->base_offset + i * ai->unit_size;
1289
1290 /* determine low/high unit_cpu */
1291 if (pcpu_low_unit_cpu == NR_CPUS ||
1292 unit_off[cpu] < unit_off[pcpu_low_unit_cpu])
1293 pcpu_low_unit_cpu = cpu;
1294 if (pcpu_high_unit_cpu == NR_CPUS ||
1295 unit_off[cpu] > unit_off[pcpu_high_unit_cpu])
1296 pcpu_high_unit_cpu = cpu;
1297 }
1298 }
1299 pcpu_nr_units = unit;
1300
1301 for_each_possible_cpu(cpu)
1302 PCPU_SETUP_BUG_ON(unit_map[cpu] == UINT_MAX);
1303
1304 /* we're done parsing the input, undefine BUG macro and dump config */
1305#undef PCPU_SETUP_BUG_ON
1306 pcpu_dump_alloc_info(KERN_DEBUG, ai);
1307
1308 pcpu_nr_groups = ai->nr_groups;
1309 pcpu_group_offsets = group_offsets;
1310 pcpu_group_sizes = group_sizes;
1311 pcpu_unit_map = unit_map;
1312 pcpu_unit_offsets = unit_off;
1313
1314 /* determine basic parameters */
1315 pcpu_unit_pages = ai->unit_size >> PAGE_SHIFT;
1316 pcpu_unit_size = pcpu_unit_pages << PAGE_SHIFT;
1317 pcpu_atom_size = ai->atom_size;
1318 pcpu_chunk_struct_size = sizeof(struct pcpu_chunk) +
1319 BITS_TO_LONGS(pcpu_unit_pages) * sizeof(unsigned long);
1320
1321 /*
1322 * Allocate chunk slots. The additional last slot is for
1323 * empty chunks.
1324 */
1325 pcpu_nr_slots = __pcpu_size_to_slot(pcpu_unit_size) + 2;
1326 pcpu_slot = memblock_virt_alloc(
1327 pcpu_nr_slots * sizeof(pcpu_slot[0]), 0);
1328 for (i = 0; i < pcpu_nr_slots; i++)
1329 INIT_LIST_HEAD(&pcpu_slot[i]);
1330
1331 /*
1332 * Initialize static chunk. If reserved_size is zero, the
1333 * static chunk covers static area + dynamic allocation area
1334 * in the first chunk. If reserved_size is not zero, it
1335 * covers static area + reserved area (mostly used for module
1336 * static percpu allocation).
1337 */
1338 schunk = memblock_virt_alloc(pcpu_chunk_struct_size, 0);
1339 INIT_LIST_HEAD(&schunk->list);
1340 schunk->base_addr = base_addr;
1341 schunk->map = smap;
1342 schunk->map_alloc = ARRAY_SIZE(smap);
1343 schunk->immutable = true;
1344 bitmap_fill(schunk->populated, pcpu_unit_pages);
1345
1346 if (ai->reserved_size) {
1347 schunk->free_size = ai->reserved_size;
1348 pcpu_reserved_chunk = schunk;
1349 pcpu_reserved_chunk_limit = ai->static_size + ai->reserved_size;
1350 } else {
1351 schunk->free_size = dyn_size;
1352 dyn_size = 0; /* dynamic area covered */
1353 }
1354 schunk->contig_hint = schunk->free_size;
1355
1356 schunk->map[0] = 1;
1357 schunk->map[1] = ai->static_size;
1358 schunk->map_used = 1;
1359 if (schunk->free_size)
1360 schunk->map[++schunk->map_used] = 1 | (ai->static_size + schunk->free_size);
1361 else
1362 schunk->map[1] |= 1;
1363
1364 /* init dynamic chunk if necessary */
1365 if (dyn_size) {
1366 dchunk = memblock_virt_alloc(pcpu_chunk_struct_size, 0);
1367 INIT_LIST_HEAD(&dchunk->list);
1368 dchunk->base_addr = base_addr;
1369 dchunk->map = dmap;
1370 dchunk->map_alloc = ARRAY_SIZE(dmap);
1371 dchunk->immutable = true;
1372 bitmap_fill(dchunk->populated, pcpu_unit_pages);
1373
1374 dchunk->contig_hint = dchunk->free_size = dyn_size;
1375 dchunk->map[0] = 1;
1376 dchunk->map[1] = pcpu_reserved_chunk_limit;
1377 dchunk->map[2] = (pcpu_reserved_chunk_limit + dchunk->free_size) | 1;
1378 dchunk->map_used = 2;
1379 }
1380
1381 /* link the first chunk in */
1382 pcpu_first_chunk = dchunk ?: schunk;
1383 pcpu_chunk_relocate(pcpu_first_chunk, -1);
1384
1385 /* we're done */
1386 pcpu_base_addr = base_addr;
1387 return 0;
1388}
1389
1390#ifdef CONFIG_SMP
1391
1392const char * const pcpu_fc_names[PCPU_FC_NR] __initconst = {
1393 [PCPU_FC_AUTO] = "auto",
1394 [PCPU_FC_EMBED] = "embed",
1395 [PCPU_FC_PAGE] = "page",
1396};
1397
1398enum pcpu_fc pcpu_chosen_fc __initdata = PCPU_FC_AUTO;
1399
1400static int __init percpu_alloc_setup(char *str)
1401{
1402 if (!str)
1403 return -EINVAL;
1404
1405 if (0)
1406 /* nada */;
1407#ifdef CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK
1408 else if (!strcmp(str, "embed"))
1409 pcpu_chosen_fc = PCPU_FC_EMBED;
1410#endif
1411#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
1412 else if (!strcmp(str, "page"))
1413 pcpu_chosen_fc = PCPU_FC_PAGE;
1414#endif
1415 else
1416 pr_warning("PERCPU: unknown allocator %s specified\n", str);
1417
1418 return 0;
1419}
1420early_param("percpu_alloc", percpu_alloc_setup);
1421
1422/*
1423 * pcpu_embed_first_chunk() is used by the generic percpu setup.
1424 * Build it if needed by the arch config or the generic setup is going
1425 * to be used.
1426 */
1427#if defined(CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK) || \
1428 !defined(CONFIG_HAVE_SETUP_PER_CPU_AREA)
1429#define BUILD_EMBED_FIRST_CHUNK
1430#endif
1431
1432/* build pcpu_page_first_chunk() iff needed by the arch config */
1433#if defined(CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK)
1434#define BUILD_PAGE_FIRST_CHUNK
1435#endif
1436
1437/* pcpu_build_alloc_info() is used by both embed and page first chunk */
1438#if defined(BUILD_EMBED_FIRST_CHUNK) || defined(BUILD_PAGE_FIRST_CHUNK)
1439/**
1440 * pcpu_build_alloc_info - build alloc_info considering distances between CPUs
1441 * @reserved_size: the size of reserved percpu area in bytes
1442 * @dyn_size: minimum free size for dynamic allocation in bytes
1443 * @atom_size: allocation atom size
1444 * @cpu_distance_fn: callback to determine distance between cpus, optional
1445 *
1446 * This function determines grouping of units, their mappings to cpus
1447 * and other parameters considering needed percpu size, allocation
1448 * atom size and distances between CPUs.
1449 *
1450 * Groups are always mutliples of atom size and CPUs which are of
1451 * LOCAL_DISTANCE both ways are grouped together and share space for
1452 * units in the same group. The returned configuration is guaranteed
1453 * to have CPUs on different nodes on different groups and >=75% usage
1454 * of allocated virtual address space.
1455 *
1456 * RETURNS:
1457 * On success, pointer to the new allocation_info is returned. On
1458 * failure, ERR_PTR value is returned.
1459 */
1460static struct pcpu_alloc_info * __init pcpu_build_alloc_info(
1461 size_t reserved_size, size_t dyn_size,
1462 size_t atom_size,
1463 pcpu_fc_cpu_distance_fn_t cpu_distance_fn)
1464{
1465 static int group_map[NR_CPUS] __initdata;
1466 static int group_cnt[NR_CPUS] __initdata;
1467 const size_t static_size = __per_cpu_end - __per_cpu_start;
1468 int nr_groups = 1, nr_units = 0;
1469 size_t size_sum, min_unit_size, alloc_size;
1470 int upa, max_upa, uninitialized_var(best_upa); /* units_per_alloc */
1471 int last_allocs, group, unit;
1472 unsigned int cpu, tcpu;
1473 struct pcpu_alloc_info *ai;
1474 unsigned int *cpu_map;
1475
1476 /* this function may be called multiple times */
1477 memset(group_map, 0, sizeof(group_map));
1478 memset(group_cnt, 0, sizeof(group_cnt));
1479
1480 /* calculate size_sum and ensure dyn_size is enough for early alloc */
1481 size_sum = PFN_ALIGN(static_size + reserved_size +
1482 max_t(size_t, dyn_size, PERCPU_DYNAMIC_EARLY_SIZE));
1483 dyn_size = size_sum - static_size - reserved_size;
1484
1485 /*
1486 * Determine min_unit_size, alloc_size and max_upa such that
1487 * alloc_size is multiple of atom_size and is the smallest
1488 * which can accommodate 4k aligned segments which are equal to
1489 * or larger than min_unit_size.
1490 */
1491 min_unit_size = max_t(size_t, size_sum, PCPU_MIN_UNIT_SIZE);
1492
1493 alloc_size = roundup(min_unit_size, atom_size);
1494 upa = alloc_size / min_unit_size;
1495 while (alloc_size % upa || ((alloc_size / upa) & ~PAGE_MASK))
1496 upa--;
1497 max_upa = upa;
1498
1499 /* group cpus according to their proximity */
1500 for_each_possible_cpu(cpu) {
1501 group = 0;
1502 next_group:
1503 for_each_possible_cpu(tcpu) {
1504 if (cpu == tcpu)
1505 break;
1506 if (group_map[tcpu] == group && cpu_distance_fn &&
1507 (cpu_distance_fn(cpu, tcpu) > LOCAL_DISTANCE ||
1508 cpu_distance_fn(tcpu, cpu) > LOCAL_DISTANCE)) {
1509 group++;
1510 nr_groups = max(nr_groups, group + 1);
1511 goto next_group;
1512 }
1513 }
1514 group_map[cpu] = group;
1515 group_cnt[group]++;
1516 }
1517
1518 /*
1519 * Expand unit size until address space usage goes over 75%
1520 * and then as much as possible without using more address
1521 * space.
1522 */
1523 last_allocs = INT_MAX;
1524 for (upa = max_upa; upa; upa--) {
1525 int allocs = 0, wasted = 0;
1526
1527 if (alloc_size % upa || ((alloc_size / upa) & ~PAGE_MASK))
1528 continue;
1529
1530 for (group = 0; group < nr_groups; group++) {
1531 int this_allocs = DIV_ROUND_UP(group_cnt[group], upa);
1532 allocs += this_allocs;
1533 wasted += this_allocs * upa - group_cnt[group];
1534 }
1535
1536 /*
1537 * Don't accept if wastage is over 1/3. The
1538 * greater-than comparison ensures upa==1 always
1539 * passes the following check.
1540 */
1541 if (wasted > num_possible_cpus() / 3)
1542 continue;
1543
1544 /* and then don't consume more memory */
1545 if (allocs > last_allocs)
1546 break;
1547 last_allocs = allocs;
1548 best_upa = upa;
1549 }
1550 upa = best_upa;
1551
1552 /* allocate and fill alloc_info */
1553 for (group = 0; group < nr_groups; group++)
1554 nr_units += roundup(group_cnt[group], upa);
1555
1556 ai = pcpu_alloc_alloc_info(nr_groups, nr_units);
1557 if (!ai)
1558 return ERR_PTR(-ENOMEM);
1559 cpu_map = ai->groups[0].cpu_map;
1560
1561 for (group = 0; group < nr_groups; group++) {
1562 ai->groups[group].cpu_map = cpu_map;
1563 cpu_map += roundup(group_cnt[group], upa);
1564 }
1565
1566 ai->static_size = static_size;
1567 ai->reserved_size = reserved_size;
1568 ai->dyn_size = dyn_size;
1569 ai->unit_size = alloc_size / upa;
1570 ai->atom_size = atom_size;
1571 ai->alloc_size = alloc_size;
1572
1573 for (group = 0, unit = 0; group_cnt[group]; group++) {
1574 struct pcpu_group_info *gi = &ai->groups[group];
1575
1576 /*
1577 * Initialize base_offset as if all groups are located
1578 * back-to-back. The caller should update this to
1579 * reflect actual allocation.
1580 */
1581 gi->base_offset = unit * ai->unit_size;
1582
1583 for_each_possible_cpu(cpu)
1584 if (group_map[cpu] == group)
1585 gi->cpu_map[gi->nr_units++] = cpu;
1586 gi->nr_units = roundup(gi->nr_units, upa);
1587 unit += gi->nr_units;
1588 }
1589 BUG_ON(unit != nr_units);
1590
1591 return ai;
1592}
1593#endif /* BUILD_EMBED_FIRST_CHUNK || BUILD_PAGE_FIRST_CHUNK */
1594
1595#if defined(BUILD_EMBED_FIRST_CHUNK)
1596/**
1597 * pcpu_embed_first_chunk - embed the first percpu chunk into bootmem
1598 * @reserved_size: the size of reserved percpu area in bytes
1599 * @dyn_size: minimum free size for dynamic allocation in bytes
1600 * @atom_size: allocation atom size
1601 * @cpu_distance_fn: callback to determine distance between cpus, optional
1602 * @alloc_fn: function to allocate percpu page
1603 * @free_fn: function to free percpu page
1604 *
1605 * This is a helper to ease setting up embedded first percpu chunk and
1606 * can be called where pcpu_setup_first_chunk() is expected.
1607 *
1608 * If this function is used to setup the first chunk, it is allocated
1609 * by calling @alloc_fn and used as-is without being mapped into
1610 * vmalloc area. Allocations are always whole multiples of @atom_size
1611 * aligned to @atom_size.
1612 *
1613 * This enables the first chunk to piggy back on the linear physical
1614 * mapping which often uses larger page size. Please note that this
1615 * can result in very sparse cpu->unit mapping on NUMA machines thus
1616 * requiring large vmalloc address space. Don't use this allocator if
1617 * vmalloc space is not orders of magnitude larger than distances
1618 * between node memory addresses (ie. 32bit NUMA machines).
1619 *
1620 * @dyn_size specifies the minimum dynamic area size.
1621 *
1622 * If the needed size is smaller than the minimum or specified unit
1623 * size, the leftover is returned using @free_fn.
1624 *
1625 * RETURNS:
1626 * 0 on success, -errno on failure.
1627 */
1628int __init pcpu_embed_first_chunk(size_t reserved_size, size_t dyn_size,
1629 size_t atom_size,
1630 pcpu_fc_cpu_distance_fn_t cpu_distance_fn,
1631 pcpu_fc_alloc_fn_t alloc_fn,
1632 pcpu_fc_free_fn_t free_fn)
1633{
1634 void *base = (void *)ULONG_MAX;
1635 void **areas = NULL;
1636 struct pcpu_alloc_info *ai;
1637 size_t size_sum, areas_size, max_distance;
1638 int group, i, rc;
1639
1640 ai = pcpu_build_alloc_info(reserved_size, dyn_size, atom_size,
1641 cpu_distance_fn);
1642 if (IS_ERR(ai))
1643 return PTR_ERR(ai);
1644
1645 size_sum = ai->static_size + ai->reserved_size + ai->dyn_size;
1646 areas_size = PFN_ALIGN(ai->nr_groups * sizeof(void *));
1647
1648 areas = memblock_virt_alloc_nopanic(areas_size, 0);
1649 if (!areas) {
1650 rc = -ENOMEM;
1651 goto out_free;
1652 }
1653
1654 /* allocate, copy and determine base address */
1655 for (group = 0; group < ai->nr_groups; group++) {
1656 struct pcpu_group_info *gi = &ai->groups[group];
1657 unsigned int cpu = NR_CPUS;
1658 void *ptr;
1659
1660 for (i = 0; i < gi->nr_units && cpu == NR_CPUS; i++)
1661 cpu = gi->cpu_map[i];
1662 BUG_ON(cpu == NR_CPUS);
1663
1664 /* allocate space for the whole group */
1665 ptr = alloc_fn(cpu, gi->nr_units * ai->unit_size, atom_size);
1666 if (!ptr) {
1667 rc = -ENOMEM;
1668 goto out_free_areas;
1669 }
1670 /* kmemleak tracks the percpu allocations separately */
1671 kmemleak_free(ptr);
1672 areas[group] = ptr;
1673
1674 base = min(ptr, base);
1675 }
1676
1677 /*
1678 * Copy data and free unused parts. This should happen after all
1679 * allocations are complete; otherwise, we may end up with
1680 * overlapping groups.
1681 */
1682 for (group = 0; group < ai->nr_groups; group++) {
1683 struct pcpu_group_info *gi = &ai->groups[group];
1684 void *ptr = areas[group];
1685
1686 for (i = 0; i < gi->nr_units; i++, ptr += ai->unit_size) {
1687 if (gi->cpu_map[i] == NR_CPUS) {
1688 /* unused unit, free whole */
1689 free_fn(ptr, ai->unit_size);
1690 continue;
1691 }
1692 /* copy and return the unused part */
1693 memcpy(ptr, __per_cpu_load, ai->static_size);
1694 free_fn(ptr + size_sum, ai->unit_size - size_sum);
1695 }
1696 }
1697
1698 /* base address is now known, determine group base offsets */
1699 max_distance = 0;
1700 for (group = 0; group < ai->nr_groups; group++) {
1701 ai->groups[group].base_offset = areas[group] - base;
1702 max_distance = max_t(size_t, max_distance,
1703 ai->groups[group].base_offset);
1704 }
1705 max_distance += ai->unit_size;
1706
1707 /* warn if maximum distance is further than 75% of vmalloc space */
1708 if (max_distance > VMALLOC_TOTAL * 3 / 4) {
1709 pr_warning("PERCPU: max_distance=0x%zx too large for vmalloc "
1710 "space 0x%lx\n", max_distance,
1711 VMALLOC_TOTAL);
1712#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
1713 /* and fail if we have fallback */
1714 rc = -EINVAL;
1715 goto out_free;
1716#endif
1717 }
1718
1719 pr_info("PERCPU: Embedded %zu pages/cpu @%p s%zu r%zu d%zu u%zu\n",
1720 PFN_DOWN(size_sum), base, ai->static_size, ai->reserved_size,
1721 ai->dyn_size, ai->unit_size);
1722
1723 rc = pcpu_setup_first_chunk(ai, base);
1724 goto out_free;
1725
1726out_free_areas:
1727 for (group = 0; group < ai->nr_groups; group++)
1728 if (areas[group])
1729 free_fn(areas[group],
1730 ai->groups[group].nr_units * ai->unit_size);
1731out_free:
1732 pcpu_free_alloc_info(ai);
1733 if (areas)
1734 memblock_free_early(__pa(areas), areas_size);
1735 return rc;
1736}
1737#endif /* BUILD_EMBED_FIRST_CHUNK */
1738
1739#ifdef BUILD_PAGE_FIRST_CHUNK
1740/**
1741 * pcpu_page_first_chunk - map the first chunk using PAGE_SIZE pages
1742 * @reserved_size: the size of reserved percpu area in bytes
1743 * @alloc_fn: function to allocate percpu page, always called with PAGE_SIZE
1744 * @free_fn: function to free percpu page, always called with PAGE_SIZE
1745 * @populate_pte_fn: function to populate pte
1746 *
1747 * This is a helper to ease setting up page-remapped first percpu
1748 * chunk and can be called where pcpu_setup_first_chunk() is expected.
1749 *
1750 * This is the basic allocator. Static percpu area is allocated
1751 * page-by-page into vmalloc area.
1752 *
1753 * RETURNS:
1754 * 0 on success, -errno on failure.
1755 */
1756int __init pcpu_page_first_chunk(size_t reserved_size,
1757 pcpu_fc_alloc_fn_t alloc_fn,
1758 pcpu_fc_free_fn_t free_fn,
1759 pcpu_fc_populate_pte_fn_t populate_pte_fn)
1760{
1761 static struct vm_struct vm;
1762 struct pcpu_alloc_info *ai;
1763 char psize_str[16];
1764 int unit_pages;
1765 size_t pages_size;
1766 struct page **pages;
1767 int unit, i, j, rc;
1768
1769 snprintf(psize_str, sizeof(psize_str), "%luK", PAGE_SIZE >> 10);
1770
1771 ai = pcpu_build_alloc_info(reserved_size, 0, PAGE_SIZE, NULL);
1772 if (IS_ERR(ai))
1773 return PTR_ERR(ai);
1774 BUG_ON(ai->nr_groups != 1);
1775 BUG_ON(ai->groups[0].nr_units != num_possible_cpus());
1776
1777 unit_pages = ai->unit_size >> PAGE_SHIFT;
1778
1779 /* unaligned allocations can't be freed, round up to page size */
1780 pages_size = PFN_ALIGN(unit_pages * num_possible_cpus() *
1781 sizeof(pages[0]));
1782 pages = memblock_virt_alloc(pages_size, 0);
1783
1784 /* allocate pages */
1785 j = 0;
1786 for (unit = 0; unit < num_possible_cpus(); unit++)
1787 for (i = 0; i < unit_pages; i++) {
1788 unsigned int cpu = ai->groups[0].cpu_map[unit];
1789 void *ptr;
1790
1791 ptr = alloc_fn(cpu, PAGE_SIZE, PAGE_SIZE);
1792 if (!ptr) {
1793 pr_warning("PERCPU: failed to allocate %s page "
1794 "for cpu%u\n", psize_str, cpu);
1795 goto enomem;
1796 }
1797 /* kmemleak tracks the percpu allocations separately */
1798 kmemleak_free(ptr);
1799 pages[j++] = virt_to_page(ptr);
1800 }
1801
1802 /* allocate vm area, map the pages and copy static data */
1803 vm.flags = VM_ALLOC;
1804 vm.size = num_possible_cpus() * ai->unit_size;
1805 vm_area_register_early(&vm, PAGE_SIZE);
1806
1807 for (unit = 0; unit < num_possible_cpus(); unit++) {
1808 unsigned long unit_addr =
1809 (unsigned long)vm.addr + unit * ai->unit_size;
1810
1811 for (i = 0; i < unit_pages; i++)
1812 populate_pte_fn(unit_addr + (i << PAGE_SHIFT));
1813
1814 /* pte already populated, the following shouldn't fail */
1815 rc = __pcpu_map_pages(unit_addr, &pages[unit * unit_pages],
1816 unit_pages);
1817 if (rc < 0)
1818 panic("failed to map percpu area, err=%d\n", rc);
1819
1820 /*
1821 * FIXME: Archs with virtual cache should flush local
1822 * cache for the linear mapping here - something
1823 * equivalent to flush_cache_vmap() on the local cpu.
1824 * flush_cache_vmap() can't be used as most supporting
1825 * data structures are not set up yet.
1826 */
1827
1828 /* copy static data */
1829 memcpy((void *)unit_addr, __per_cpu_load, ai->static_size);
1830 }
1831
1832 /* we're ready, commit */
1833 pr_info("PERCPU: %d %s pages/cpu @%p s%zu r%zu d%zu\n",
1834 unit_pages, psize_str, vm.addr, ai->static_size,
1835 ai->reserved_size, ai->dyn_size);
1836
1837 rc = pcpu_setup_first_chunk(ai, vm.addr);
1838 goto out_free_ar;
1839
1840enomem:
1841 while (--j >= 0)
1842 free_fn(page_address(pages[j]), PAGE_SIZE);
1843 rc = -ENOMEM;
1844out_free_ar:
1845 memblock_free_early(__pa(pages), pages_size);
1846 pcpu_free_alloc_info(ai);
1847 return rc;
1848}
1849#endif /* BUILD_PAGE_FIRST_CHUNK */
1850
1851#ifndef CONFIG_HAVE_SETUP_PER_CPU_AREA
1852/*
1853 * Generic SMP percpu area setup.
1854 *
1855 * The embedding helper is used because its behavior closely resembles
1856 * the original non-dynamic generic percpu area setup. This is
1857 * important because many archs have addressing restrictions and might
1858 * fail if the percpu area is located far away from the previous
1859 * location. As an added bonus, in non-NUMA cases, embedding is
1860 * generally a good idea TLB-wise because percpu area can piggy back
1861 * on the physical linear memory mapping which uses large page
1862 * mappings on applicable archs.
1863 */
1864unsigned long __per_cpu_offset[NR_CPUS] __read_mostly;
1865EXPORT_SYMBOL(__per_cpu_offset);
1866
1867static void * __init pcpu_dfl_fc_alloc(unsigned int cpu, size_t size,
1868 size_t align)
1869{
1870 return memblock_virt_alloc_from_nopanic(
1871 size, align, __pa(MAX_DMA_ADDRESS));
1872}
1873
1874static void __init pcpu_dfl_fc_free(void *ptr, size_t size)
1875{
1876 memblock_free_early(__pa(ptr), size);
1877}
1878
1879void __init setup_per_cpu_areas(void)
1880{
1881 unsigned long delta;
1882 unsigned int cpu;
1883 int rc;
1884
1885 /*
1886 * Always reserve area for module percpu variables. That's
1887 * what the legacy allocator did.
1888 */
1889 rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE,
1890 PERCPU_DYNAMIC_RESERVE, PAGE_SIZE, NULL,
1891 pcpu_dfl_fc_alloc, pcpu_dfl_fc_free);
1892 if (rc < 0)
1893 panic("Failed to initialize percpu areas.");
1894
1895 delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
1896 for_each_possible_cpu(cpu)
1897 __per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu];
1898}
1899#endif /* CONFIG_HAVE_SETUP_PER_CPU_AREA */
1900
1901#else /* CONFIG_SMP */
1902
1903/*
1904 * UP percpu area setup.
1905 *
1906 * UP always uses km-based percpu allocator with identity mapping.
1907 * Static percpu variables are indistinguishable from the usual static
1908 * variables and don't require any special preparation.
1909 */
1910void __init setup_per_cpu_areas(void)
1911{
1912 const size_t unit_size =
1913 roundup_pow_of_two(max_t(size_t, PCPU_MIN_UNIT_SIZE,
1914 PERCPU_DYNAMIC_RESERVE));
1915 struct pcpu_alloc_info *ai;
1916 void *fc;
1917
1918 ai = pcpu_alloc_alloc_info(1, 1);
1919 fc = memblock_virt_alloc_from_nopanic(unit_size,
1920 PAGE_SIZE,
1921 __pa(MAX_DMA_ADDRESS));
1922 if (!ai || !fc)
1923 panic("Failed to allocate memory for percpu areas.");
1924 /* kmemleak tracks the percpu allocations separately */
1925 kmemleak_free(fc);
1926
1927 ai->dyn_size = unit_size;
1928 ai->unit_size = unit_size;
1929 ai->atom_size = unit_size;
1930 ai->alloc_size = unit_size;
1931 ai->groups[0].nr_units = 1;
1932 ai->groups[0].cpu_map[0] = 0;
1933
1934 if (pcpu_setup_first_chunk(ai, fc) < 0)
1935 panic("Failed to initialize percpu areas.");
1936}
1937
1938#endif /* CONFIG_SMP */
1939
1940/*
1941 * First and reserved chunks are initialized with temporary allocation
1942 * map in initdata so that they can be used before slab is online.
1943 * This function is called after slab is brought up and replaces those
1944 * with properly allocated maps.
1945 */
1946void __init percpu_init_late(void)
1947{
1948 struct pcpu_chunk *target_chunks[] =
1949 { pcpu_first_chunk, pcpu_reserved_chunk, NULL };
1950 struct pcpu_chunk *chunk;
1951 unsigned long flags;
1952 int i;
1953
1954 for (i = 0; (chunk = target_chunks[i]); i++) {
1955 int *map;
1956 const size_t size = PERCPU_DYNAMIC_EARLY_SLOTS * sizeof(map[0]);
1957
1958 BUILD_BUG_ON(size > PAGE_SIZE);
1959
1960 map = pcpu_mem_zalloc(size);
1961 BUG_ON(!map);
1962
1963 spin_lock_irqsave(&pcpu_lock, flags);
1964 memcpy(map, chunk->map, size);
1965 chunk->map = map;
1966 spin_unlock_irqrestore(&pcpu_lock, flags);
1967 }
1968}